News
What's happening in robotics and AI — curated by the wujec.ai editors.
Yukai built its name on robots that hear nothing. Its family robot now remembers your conversations for 1,700 yen a month
On 28 August 2026 Yukai Engineering switched on a generative-AI conversation mode in BOCCO emo, the family robot it has sold since March 2021, and put it behind a paid subscription: a premium plan at 1,700 yen a month, with two months free for anyone signing up before the end of September 2026. What the money buys is memory. In the new mode the robot recognises which member of the household is speaking, remembers what was said and carries that context from one conversation into the next, rather than treating each exchange as a fresh voice command. The same character also appears as an illustration inside the app's chat room, so a conversation can continue on a phone away from home. A free update landed the same day for every owner, and it is the more revealing half of the announcement. Until now the robot had to be summoned by name; it can now be woken by poking it or by pressing its function button. The recording window, previously six seconds, was extended to ten. Both changes read as fixes for a device whose voice interface was tuned for short instructions and now has to cope with something closer to talking. The move is worth noting because of what else this manufacturer sells. Yukai's best-known products are deliberately deaf: Qoobo, the tailed cushion, has no microphone and no camera at all, which is exactly why it is accepted in care homes and hospital wards. BOCCO emo has always been the opposite case in the same catalogue — mains-powered, cloud-dependent, useless without a router and a smartphone — and the premium plan pushes it further in that direction. Yukai has not said which AI model or provider sits behind the conversation feature; its announcement says only that it uses generative AI. The Wi-Fi model of BOCCO emo currently lists at 52,800 yen, up from the 44,000 yen it launched at in 2021. Buyers who want the new conversations should count on the subscription on top.
BOCCO emo →For seven years Kassow left repeatability off every spec sheet. Its two new arms finally print a number
Kassow Robots has spent its whole commercial life as the odd one out in the cobot market. Every arm it builds has seven axes rather than six, and until this spring not one of them had a published repeatability figure - not in the product manual, not in the datasheet section, not in the FAQ. In a class where every competitor prints a tolerance in hundredths of a millimetre, the Danish maker printed nothing. That changed with the KR1240 and KR1824, introduced on 10 April 2026. The brochure for the pair lists repeatability of +/-0.03 mm for the KR1240 and +/-0.05 mm for the KR1824 - ordinary numbers for this class, but the first the company has ever committed to. The five older arms, KR810 through KR1805, still carry no figure. The two models also mark the first real mechanical break in the range. The KR1240 carries 40 kg over 1200 mm and the KR1824 carries 24 kg over 1800 mm, against 5 to 18 kg on the earlier arms. Kassow attributes the jump to new joints: 1.5 times the torque, 40 percent more mechanical stiffness and wrist joints roughly 20 percent faster, reaching 270 deg/s on the KR1824. The KR1824 is the more telling of the pair. Its 1800 mm reach is not new - the KR1805 has covered that radius since 2019 - but it did so with payload cut to 5 kg. The new arm keeps the radius and lifts nearly five times as much. The price is size. The base grows from 160 x 160 mm to 194 x 194 mm, and mass roughly doubles: 80 kg for the KR1240 and 76 kg for the KR1824, against 34 to 38 kg for the older models. Protection improves to IP56 and IP55 respectively, up from IP54. Both are offered in the Edge Edition, with the controller folded into the arm base for mobile platforms. Both models now have profiles in our catalogue, and the five older entries have been corrected: the claim that Kassow publishes no repeatability for any model was true until April 2026 and is not any more.
Kassow Robots KR1240 →Elite Robots' three biggest cobots weigh the same. Their payload climbs from 20 to 30 kg
Elite Robots of Suzhou sells six collaborative arms in its CS series, and the numbers on its own product pages describe a range that stops growing physically while it keeps growing in capability. The first three models scale the way arms usually do. The CS63 carries 3 kg and weighs 15 kg; the CS66 carries 6 kg and weighs 20 kg; the CS612 carries 12 kg and weighs 34 kg. Payload roughly doubles, mass rises with it, and the ratio of what the machine lifts to what it weighs improves from 0.20 to 0.35. At the top of the range the mass stops. The CS620 lifts 20 kg and weighs 60 kg. The CS625 lifts 25 kg and weighs 58 kg. The CS530H, the newest and strongest arm in the family, lifts 30 kg and also weighs 58 kg. Half a family's worth of extra payload arrives on a body that never gets heavier — the payload-to-weight ratio rises from 0.33 to 0.52, and the strongest arm is fractionally the lightest of the three. What pays for it is reach. The CS620 stretches to 1,800 mm, the longest arm the company builds; the CS625 pulls back to 1,500 mm and the CS530H to 1,522 mm. Shortening the lever is what lets the same motors hold more weight. The same trade shows up in accuracy: the CS625, on the shortest of the three arms, is the most repeatable at +/-0.08 mm, while both 30 kg and 20 kg models sit at +/-0.1 mm. The CS530H also breaks a habit of the category. Its tool point moves at 4.0 m/s, the fastest in the family, while carrying the heaviest load — the CS625 manages 3.3 m/s with 5 kg less. Elite Robots describes the model as built around palletising, reaching stacking heights of up to 2,300 mm, which explains a design tuned for vertical throughput rather than radius. Across the whole six-model range the pattern of the industry holds: precision degrades as the arm gets longer, from +/-0.02 mm on the 624 mm CS63 to +/-0.1 mm on the 1,800 mm CS620 — a fivefold spread. It is the same relationship visible in Universal Robots' range, where the smallest and oldest arm is still the most accurate. The manufacturer does not date the CS530H publicly. Web archive snapshots place it: the company's product listing of 19 June 2024 already showed the other five CS arms and not this one, and its product page has no snapshot before 19 March 2025. The three smallest models appear in the listing archived on 31 March 2023, consistent with the manufacturer's own statement that the CS series was released in 2022.
Elite Robots CS530H →Dobot quietly grew its cobot range to ten arms — and the longest one is also the most precise
Dobot's CRA specification table listed seven collaborative arms in August 2025. It lists ten today. The three additions — CR6AL, CR16AL and CR30A — arrived without a launch announcement, and one of them quietly contradicts a rule this catalogue has watched hold at four other arm makers. The new CR6AL reaches 1,800 mm, further than any other arm Dobot sells and 100 mm beyond the 20 kg flagship. Reach normally costs precision: a longer lever amplifies every error at the joints. Yet Dobot rates the CR6AL at ±0.02 mm repeatability, the tightest figure in the whole family, while the CR7A — which reaches 800 mm, less than half as far — is rated at ±0.05 mm, two and a half times looser. Within a single manufacturer's own table, the longest arm is the most accurate and one of the shortest is the least. The payload numbers refuse to line up too. The CR6AL lifts 6 kg, less than the shorter CR7A's 7 kg. The CR5A reaches 900 mm while the stronger CR7A reaches 800 mm. And the new CR30A, the strongest arm Dobot has ever sold at 30 kg, covers a radius of 1,305 mm — 395 mm short of the 20 kg CR20A — at 4 m/s rather than 5.5 m/s. It is nonetheless three kilograms lighter than the CR20A and runs on the same 500 W supply, so the extra payload was bought with reach and speed, not with more metal. Dating the three arms took the Internet Archive and the web server. They are absent from a snapshot of 11 January 2026 and present on 19 July 2026; their product images carry a Last-Modified header of 20 March 2026, all three uploaded in the same second. This catalogue therefore dates them to 2026 and says in each profile that the year comes from snapshots rather than from a manufacturer's statement. Two inconsistencies in Dobot's published data are worth recording. The series table gives maximum linear speeds of 3 to 5.5 m/s, while the CR20A's own product page puts maximum tool-centre-point speed at 2,000 mm/s — a gap of nearly threefold with no explanation. And the ingress-protection row of the series table shows a repeatability value instead of an IP rating for the two largest models; the CR20A page states IP54 for arm and control box, which is the figure published here. Six of the ten arms now have profiles on wujec.ai: CR3A, CR5A, CR7A, CR6AL, CR20A and CR30A. The remaining four will follow.
Dobot CR6AL →Flexiv's two small Rizon arms lost 40 mm of reach in the autumn of 2024 — and the Rizon 10's claimed touch got twice as coarse
Flexiv's Rizon series specification table was rewritten between 3 August and 9 October 2024, and none of the model names changed with it. On the earlier version of the page the Rizon 4 reached 916 mm and the Rizon 4s 959 mm; on the later one, and on the page as it stands today, they reach 876 mm and 919 mm. Both arms lost exactly 40 mm. In the same edit the Rizon 10's force sensing accuracy moved from 0.1 N to 0.2 N — a declaration made twice as coarse, on a machine sold on its sense of touch. Everything else in those columns stayed put: payload 4, 4 and 10 kg, arm weight 20, 21 and 38 kg, pose repeatability ±0.05 mm to ISO 9283, IP65, maximum tool force 200, 150 and 350 N. The Rizon 10's own reach, 941 mm, did not move either. The change was not announced. Flexiv's company timeline records product launches, awards, funding rounds and certifications, but carries nothing for the autumn of 2024 beyond the release of Flexiv Elements Series 3 in October. There is no note on the product page, and the arms are still sold under the same names, which means a datasheet printed in mid-2024 and one printed today describe the same robot with different numbers. Which set is right matters practically: 40 mm is roughly the difference between reaching the far corner of a work cell and not reaching it. The current table is the manufacturer's own, so this catalogue uses 876 mm and 919 mm — and the correction also explains a long-running disagreement, because several distributor listings still credit the Rizon 4 with 780 mm and the Rizon 10 with 1015 mm, matching neither version of the manufacturer's page. The same archived pages date one more thing Flexiv does not: the Rizon 10s, the fourth model in the family, first appears in the specification table between 14 June and 5 August 2025. It has no launch announcement and no entry on the company timeline.
Flexiv Rizon 4 →Read JAKA's Zu30 page in inches and pounds and you are reading the Zu20
Every headline figure on the product page of JAKA's 30 kg collaborative arm is printed twice, once in metric and once in imperial. The two do not describe the same machine. Converted, the imperial column turns out to be the specification of the neighbouring model, the Zu20. The page states a payload of "30 kg (44.1 lb)". Thirty kilograms is 66.1 lb; 44.1 lb is 20.0 kg. It states a reach of "1350 mm (70.1 in)". Seventy-point-one inches is 1,781 mm. It states a weight of "65 kg (149.9 lb)". One hundred and forty-nine-point-nine pounds is 68.0 kg. Payload 20 kg, reach 1,780 mm, weight 68 kg: those are, to the digit, the three numbers JAKA publishes for the Zu20 one page across. The imperial column was not mistyped — it was copied wholesale from a different robot. The practical consequence falls on exactly the buyers the units are there for. A North American integrator sizing a cell in inches and pounds from this page will plan around an arm that reaches 430 mm farther and lifts a third less than the one being ordered — and both errors point in the direction that produces an unusable workcell rather than a merely tight one. The error is not isolated. Across JAKA's classic Zu family the same parameter table is shifted by one row, so the field labelled "Degree of freedom" displays a programming method and a field labelled "Humidity" displays the tool input/output count. On the desktop MiniCobo the same shift pushes the arm's ingress protection rating off the table altogether: the field named "IP classification" shows an ambient temperature range, and the actual IP figure is nowhere on the page. The company's newest tables — the Zu35 and the Mini 2 — carry the same fields correctly labelled, which suggests an old template still serving most of the catalogue rather than a one-off slip. None of this touches the robots themselves. JAKA, founded in Shanghai in 2014 and the manufacturer Teradyne is currently suing at Europe's Unified Patent Court, builds a well-regarded range whose metric figures are internally consistent: within the classic Zu family repeatability loosens from ±0.02 mm on the short arms to ±0.05 mm on the long ones, exactly as arm length predicts. It is the sheet that is broken, not the machine. Three JAKA arms — the Zu3, the Zu12 and the MiniCobo — enter this catalogue today, with metric values only and with the shifted rows left blank rather than filled in. Figures verified against JAKA's own English product pages on 8 September 2026.
JAKA Zu12 →Techman has launched its smallest cobot ever — and left it out of its own product catalogue
Techman Robot put a new collaborative arm on the market on 24 July 2026, and six weeks later the machine still has no page in the company's product catalogue. The TM3S is the smallest arm the Taiwanese manufacturer has ever sold: 3 kg of payload, a 600 mm working radius and a body it describes as weighing under 12 kg. The number that matters is the last one. At under 12 kg the arm is inside what one technician can lift off a bench and carry, and Techman designed the rest of the machine around that possibility. The power lead uses a quick-release connector, so the composite cable can be swapped and the unit unbolted and remounted at another station without opening the arm. On a shop floor running short batches, that turns relocation from a scheduled job into something done between shifts. Everything else is borrowed from the S series launched in 2023: the same control box, the same TMflow programming environment, the same hand-guided vision teaching, and the wrist camera that has been standard on every Techman arm since 2016. The manufacturer points the TM3S at precision welding, assembly, workstation localisation and fast automated inspection in confined spaces. What is missing is the paperwork. The arm made its overseas debut in Bangkok at ME Assembly and Automation on 17 June 2026 and was declared commercially available five weeks later, but as of 8 September 2026 tm-robot.com has no TM3S entry under either cobot series, no datasheet and no price. The payload, reach and mass quoted above exist in exactly one place: a launch post on the company's Chinese-language site. Figures the S series is otherwise sold on — repeatability of plus/minus 0.03 mm, IP65 sealing — have not been published for this model, and this catalogue does not assume they carry over. The launch also settles a question about the rest of the range. Techman has added a ninth S-series model without dropping anything: TM5-700, TM5-900, TM12, TM14, TM16 and TM20, all from 2016 to 2021, remain listed and linked from the manufacturer's front page. Where Universal Robots renamed two of its best-known arms this year and FANUC quietly cut its collaborative line from five models to two, Techman is still adding to both ends at once.
Techman TM3S →Universal Robots' two best-known cobots gained 2.5 kg each without a single new part
Anyone looking today for a UR5e or a UR10e on the maker's website ends up on the general product listing. Both pages redirect, while the UR3e and UR16e pages still answer at their own addresses — which rules out a site rebuild and points at something else. The two most widely installed collaborative arms in the world left the catalogue under those names on 12 May 2025. They did not leave the factory. At the Automate fair in Detroit, alongside the launch of the faster UR15, Universal Robots announced that the UR5e would be sold as the UR7e and the UR10e as the UR12e. The company's own wording was that it was introducing clearer names and more power where it counts. Deliveries of both were declared immediate, and the e-Series line-up became UR3e, UR7e, UR12e and UR16e. Put the old and the new technical data sheets side by side and the interesting part appears: on the mechanical side, nothing moved. The UR7e weighs 20.6 kg with its cable, reaches 850 mm, stands on a base 151 mm across, repeats a pose to within 0.03 mm per ISO 9283, tops out at 4 m/s at the tool centre point and draws 570 W at maximum — every one of those figures matches the UR5e document. The UR12e repeats the UR10e in the same way: 33.5 kg, 1,300 mm, a 190 mm base, 0.05 mm of repeatability, 615 W, and a force-torque sensor rated to 100 N and 10 Nm. Universal Robots states that accessories, URCaps and application kits transfer unchanged and that installed systems need no modification. What changed is the label. The smaller arm now carries a 7.5 kg rating instead of 5 kg — an increase of half again on a machine whose parts are the same. The larger one gained nothing at all in 2025: it had been rated at 12.5 kg since an enhanced version announced on 17 May 2021, sold at an unchanged price, so for four years its name understated its lifting capacity by a fifth. The 2025 announcement corrected a number that had been wrong since then. For a buyer the practical consequence is small but real. A quotation, a risk assessment or an integrator's cell drawing dated before May 2025 will name a robot that no longer appears in any price list, and a specification written around a 5 kg limit is being met by an arm the maker now certifies for 7.5 kg. For a catalogue like this one the consequence is a rule worth remembering: a model vanishing from a manufacturer's menu does not always mean the machine was discontinued. Sometimes it means only that the marketing department finally read the data sheet.
UR7e →FANUC's green cobot line has quietly shrunk from five arms to two
FANUC's collaborative robots come in two colours. The white CRX arms are the ones the company advertises. The green CR arms came first, and there are now only two of them left on sale. A reading of the manufacturer's current cobot listing shows eight models: the CRX-3iA, CRX-5iA, CRX-10iA, CRX-10iA/L, CRX-20iA/L and CRX-30iA in white, and just the CR-15iA and CR-35iB in green. Missing are the CR-4iA, CR-7iA and CR-7iA/L that opened the line at IMTS in September 2016, and the CR-35iA that started it all in June 2015. None of the four has been declared discontinued. Their product pages are still published and still carry full specification tables — but where a current model offers a "Get a Quote" button, each of the four now offers "Get Support". That single word is the only public statement of their status, and it is the reason this catalogue has moved all four to retired without a withdrawal date: FANUC has published none. The replacement tells its own story. The CR-35iA of 2015 was the first force-limited arm to carry more than 10 kg, and it needed 990 kg of machine to lift 35 kg — twenty-eight kilograms of robot per kilogram of payload. The CR-35iB that replaced it in 2022 lifts 50 kg, since a 2023 uprating, using 375 kg of robot. Seven years bought a higher payload for 615 kg less mass, at the cost of 170 mm of reach. The trade the buyer accepted in 2015 — enormous mass for the right to work without a fence — is no longer the trade on offer. This catalogue keeps profiles for all six green arms, including the four no longer sold. A machine that defined a market segment does not stop being worth describing when its maker stops quoting a price for it.
CR-35iA →The name says 35, the datasheet says 50: FANUC's strongest cobot outgrew its own model number
FANUC's CR-35iB is, by the manufacturer's own description, the industry's strongest collaborative robot: 50 kg of payload with no safety fence, over a 1,643 mm reach, from an arm weighing 375 kg. The figure that should give a specifier pause is the one in the name. The machine launched in May 2022 as a 35 kg cobot, a redesign of the 2015 CR-35iA that shed roughly 62 percent of its mass, 375 kg against 990 kg, to make the heavy cobot something a plant could install in an afternoon. On 1 June 2023, after showing the robot at Automate in Detroit, FANUC America announced a payload increase to 50 kg. The model number did not change. FANUC's own cobot microsite still describes the machine as having "a payload of 35kg or 50kg", while its specification table lists 50 kg flat. The same announcement did the opposite to a sister machine. The CRX-25iA, FANUC's largest white-series cobot, also gained payload - from 25 kg to 30 kg - and this one was renamed, becoming the CRX-30iA. FANUC's website now redirects the old name to the new page. Two upgrades announced on the same day, two different conventions: one arm kept a number that understates it by 15 kg, the other took a new one. For anyone reading a specification sheet this is a practical trap rather than a curiosity. A 2022 datasheet, a 2026 quotation and a third-party listing can all carry the designation CR-35iB and describe different lifting capacities, and none of them is wrong. FANUC publishes no list price for either machine, so the payload line in a written quotation is the only place the difference becomes visible. This catalogue records the manufacturer's current specification table - 50 kg - and notes the older 35 kg rating in the profile, because both appear to remain in the manufacturer's own materials. Two further discrepancies sit in the same file: FANUC America gives the reach as 1,643 mm while the 2022 European launch release quoted 1,831 mm, and the product page lists 300 mm/s in collaborative mode where the cobot microsite lists 750 mm/s.
CR-35iB →Two FANUC cobots, one 1,418 mm arm: the stronger one is the only cobot in the range that never speeds up
FANUC sells two collaborative arms with exactly the same 1,418 mm reach, and the specification tables on its own product pages show the stronger one giving up something the rest of the range keeps: its high-speed mode. The CRX-10iA/L, launched in December 2019, carries 10 kg and weighs 39 kg. The CRX-20iA/L, announced on 1 March 2022 and shipping from that April, carries 20 kg and weighs 41 kg. Same reach, same six axes, same R-30iB Plus controller, same IP67 sealing, same mounting options. Twice the payload for two kilograms of extra arm, and the heavier model is the more accurate of the two at +/-0.04 mm against +/-0.05 mm. On every line a buyer normally reads, the 20 kg arm is simply the better machine. The exception is the speed row. Every other arm FANUC lists in the CRX family runs at 1,000 mm/s in collaborative mode and doubles to 2,000 mm/s in high-speed mode, the setting used when the cell is fenced and no operator is standing inside it. For the CRX-20iA/L, FANUC publishes 1,000 mm/s in both modes. The strongest mid-size cobot in the range is capped at human-safe speed permanently, so in a fenced cycle the older, weaker CRX-10iA/L will complete more moves per hour on any job it can physically lift. That matters because most cobots do not spend their lives working next to people. A great many end up behind a fence doing palletizing or machine tending, where the collaborative rating is an installation convenience rather than a daily operating mode. For those cells the missing high-speed mode is a throughput ceiling, and it is not visible in the payload, the reach or the price. The same tables settle a question this catalogue has now asked of four cobot makers in a row. Reach does not predict payload. Techman numbers its arms so that a TM5-700 lifts more than a TM5-900; Universal Robots gives its strongest and its mid-range arms an identical reach; KUKA and now FANUC repeat the pattern. In FANUC's case two arms of the same length differ by a factor of two in what they will carry. The number in the name and the number on the drawing are not the same kind of fact, and only one of them is in the datasheet. wujec.ai has added the CRX-10iA, CRX-10iA/L, CRX-20iA/L and CRX-30iA to the catalogue. All figures come from FANUC America's own product specification tables.
CRX-20iA/L →Techman's S series is not a new cobot line. It is the old arms, rebuilt to a third of the tolerance
Techman Robot unveiled the TM AI Cobot S series at Automate 2023 in Detroit on 22 May 2023. Read the four base models against the arms they replace, and the series turns out to be a rebuild rather than a new range — the geometry is carried over almost unchanged, and what moves is the precision. The TM12S carries 12 kg over 1300 mm and weighs 33 kg. The TM12 of 2018 carried 12 kg over 1300 mm and weighed 32.8 kg. The difference is repeatability: plus/minus 0.1 mm then, plus/minus 0.03 mm now. The TM14S repeats the pattern against the TM14 — same 14 kg, same 1100 mm, tolerance cut by the same factor of a little over three. That is where Techman's advertised "up to 70 percent" improvement in repeatability is literally true. At the top of the range it is not: the company's own product pages give the TM25S and TM30S plus/minus 0.05 mm. The smaller pair moved further. The TM5S keeps the 946 mm reach of the 2016 TM5-900 but lifts 5 kg instead of 4, and the TM7S takes the place of the TM5-700 with 7 kg over 758 mm instead of 6 kg over 700 mm. Both gained the same tolerance step, from plus/minus 0.05 mm to plus/minus 0.03 mm. The rest of the S-series package is what a 2016 or 2018 arm does not have: IP65 sealing instead of IP54, so the machine can be hosed down; a sixth axis turning at 450 rather than 225 degrees per second, which the manufacturer converts into a claimed 25 percent shorter cycle; and up to 31 TUV-certified safety functions rated PL d, Cat 3, which Techman says remove the need for a separate safety PLC in the cell. One detail is worth reading carefully before ordering. In the older line the model number named the family and a suffix gave the reach, so a TM5-700 was the short version of a TM5. In the S series the number is the payload in kilograms and the reach is not in the name at all — which is why the TM5S reaches 188 mm further than the TM7S sitting next to it in the same catalogue. The catalogue now holds all eight S-series arms alongside the older line they were built from.
Techman TM12S →Techman sells its S series on +/-0.03 mm. Its own pages give the two biggest arms +/-0.05
Techman Robot promotes its TM AI Cobot S series on one number: a repeatability of plus/minus 0.03 mm, which the company calls an improvement of up to 70 percent over its earlier line. The claim sits on the series landing page, and trade coverage of the TM25S and TM30S launches repeated it as a specification of those machines. The manufacturer's own product pages say something else. Read one by one, the eight arms of the S series split into two groups. TM5S, TM7S, TM12S, TM14S and TM20S each carry the plus/minus 0.03 mm line. The TM25S and the TM30S — the 25 kg arm with the longest reach in the catalogue at 1902 mm, and the 30 kg arm rated to 35 kg when palletising — are both listed at plus/minus 0.05 mm. The practical reading is that the series headline describes the light and medium arms, not the flagship pair. A buyer who needs both the reach and the precision has to choose: the TM20S holds plus/minus 0.03 mm but stops at 1300 mm and 20 kg, while going up to the TM25S buys 600 mm of extra span at a stated cost of two hundredths of a millimetre. The difference is small in absolute terms and irrelevant for palletising, which is what these two arms are sold for. It is not irrelevant for the buyer comparing spreadsheets, because the number quoted in most launch coverage is not the number on the product page. Where the two sources disagree, this catalogue records the per-model figure from the manufacturer's current page and says so in the profile. wujec.ai has added the TM20S, TM25S and TM30S to the catalogue today, completing Techman's collaborative range at eight S-series arms and seven of the earlier line.
Techman TM25S →Two cobots, the same 900 mm arm, twice the difference in precision — and the reason is not the reach
Last week the Universal Robots range gave us a tidy rule: pose repeatability degrades as the arm gets longer, almost in proportion. Put the Techman Robot catalogue next to it and the rule falls apart. Techman's TM5-900 and its TM16 both work within roughly a 900 mm radius. The TM5-900 is quoted at +/-0.05 mm repeatability, the TM16 at +/-0.1 mm. Same reach, half the precision. The number that changed is the payload: 4 kg against 16 kg. And once a Techman arm leaves the TM5 series, the tolerance stops moving at all — the 900 mm TM16, the 1,100 mm TM14, the 1,300 mm TM12 and the 1,300 mm TM20 are every one of them rated at +/-0.1 mm, whatever their length. So precision at this manufacturer is a property of the platform, not of the geometry. There are two families: a light one that holds 50 micrometres and a heavy one that holds 100, and choosing a shorter model inside the heavy family buys nothing in accuracy. That is a different engineering answer to the same physics, and it changes how a cell is specified. At Universal Robots a buyer trades reach for precision. At Techman a buyer trades payload class for precision, and the reach comes free. The pattern repeats inside the heavy family in the way speed is spent. The TM12 and the TM20 share a 1,300 mm arm and weigh 32.8 and 33 kg — two hundred grammes apart — but the TM20 carries eight kilogrammes more. It pays for them with time rather than metal: its first two joints turn at 90 degrees per second where the TM12's run at 120, and average tool speed falls from 1.3 to 1.1 metres per second. The TM12 is, quietly, the fastest arm the company sells, because a long arm converts a given joint rate into more travel at the tool. One more number is worth pulling out, because it runs against the industry grain. Every arm in this range is rated for an ISO Class 3 cleanroom, a stricter class than the Class 5 that only a single Universal Robots model reaches and the Class 6 available on its larger ones. For semiconductor and optics work — where the whole question is whether the machine may stand in the clean space at all — that is the line in the specification that decides the purchase, and here it does not depend on which model is chosen. A note on where the figures come from. Techman publishes no specification table on its own product pages, and its printed catalogue carries the numbers as an image with no text layer. The full data set is documented instead by Omron, which has resold these arms under its own brand since the two firms allied in 2018; its comparison table was last revised in October 2024. Where the two sources disagree — Omron rounds the TM5 reaches to 700 and 900 mm, Techman states 946 mm for the long variant — the manufacturer's own figure stands.
Techman TM16 →The most precise arm Universal Robots sells is the smallest one it launched in 2018
Put the whole Universal Robots range side by side and one number refuses to improve with time. Pose repeatability, measured per ISO 9283, is +/-0.03 mm on the UR3e and the UR5e — the two smallest arms of the e-Series, shown at automatica in June 2018. Every machine the company has launched since is rated at +/-0.05 mm or worse: the UR16e of 2019 and the UR10e at +/-0.05 mm, the UR15 of 2025 at +/-0.05 mm, and the heavy UR20 at +/-0.1 mm, three times the figure of an arm now eight years old. The pattern is not a decline in engineering. It is what reach and payload cost. Longer tubes and heavier joints move a fixed positional error further from the base, and in this family the relationship is almost mechanical: the 500 mm UR3e and 850 mm UR5e hold 30 micrometres, the 900 mm UR16e and 1,300 mm UR10e hold 50, and the 1,750 mm UR20 holds 100. What makes it worth stating plainly is that the newest products do not escape it. The UR15, introduced in 2025 as the company's fastest arm, matches the 2018 UR10e on repeatability while beating it on speed. Universal Robots has spent seven years adding payload, reach and tool speed; the accuracy ceiling of the small arms has not moved. For a buyer, that inverts the usual instinct to specify the newest and largest machine that fits the budget. An application defined by where the part is placed — electronics assembly, laboratory handling, screwdriving, dispensing — is best served by the oldest and cheapest arms in the catalogue. Reach is what you buy at the top of the range; precision is what you buy at the bottom. Two smaller findings sit in the same data sheets. The UR3e is the only arm in the family whose three wrist joints run at 360 degrees per second, twice the rate of everything else, which suits it to twisting work rather than carrying. And the UR5e is the only Universal Robots arm certified for an ISO Class 5 cleanroom; the larger UR10e and UR16e reach only Class 6. In this range, cleanliness and precision both go down as the arm goes up. All figures are from the manufacturer's own technical data sheets, revision of May 2025, and its PolyScope 5 user manual.
UR3e →The company that invented the cobot is now defending it in court — across 17 countries at once
On 27 August 2026 Teradyne Robotics A/S filed a patent infringement case at the Unified Patent Court in Copenhagen against the German subsidiary of the Chinese manufacturer JAKA. The patents at issue are both software and hardware, and belong to the business unit that created this market: Universal Robots of Odense, whose UR5 established the collaborative robot as a product category in 2008. The venue is the point. The Unified Patent Court is a single court whose rulings take effect in 17 of the 18 EU member states that take part in it. A patent dispute that would once have been fought country by country can now settle the European availability of a whole product line in one judgment. Teradyne says the case covers a broad range of JAKA cobot models sold in the EU. It is the company's second intellectual property action in Europe this year. The earlier one ran in Germany against a subsidiary of Elite Robots and concerned copyright in Universal Robots' own software; the German court issued an injunction in Teradyne's favour. JAKA rejects the claim. The company says its technology comes from years of its own research, that it holds more than 300 granted patents worldwide, and that it commissioned two independent freedom-to-operate analyses before entering the European market. Why this matters beyond two company names: collaborative arms from Chinese manufacturers now undercut European ones sharply on price, and the incumbent's answer is being tested in a court whose reach is continental rather than national. No ruling has been issued, and nothing here is a finding against JAKA — the case has only been filed.
UR20 →In KUKA's cobot range, the cleaner the room, the bigger the robot you have to buy
KUKA now lists seven builds of its LBR iisy collaborative arm, and two of them are factory cleanroom versions rated up to ISO Class 3. Both sit at the top of the payload range: the 11 kg long-arm model and, since 2025, the 15 kg compact one. The 3, 6 and 8 kg arms have no cleanroom variant at all. That ordering runs against the work. Cleanroom duty is concentrated in semiconductor handling, optics, medical device assembly and laboratory automation, and those are places where benches are short, equipment is packed tightly and a small arm is usually the right tool. In this family a buyer who needs ISO Class 3 has no small option: the entry point is a 1300 mm arm rated for 11 kg. The same inversion shows at the other end of the range. The smallest LBR iisy, the 3 kg R760, is the only member rated IP30 — protection against solid bodies above 2.5 mm and no protection against water whatsoever. Every model from 6 kg upward is IP54. In the family table the row looks like any other, but the practical meaning is a hard boundary: the 3 kg arm belongs on an assembly bench, not in a machining cell with coolant mist, grinding dust or wash-down cleaning. So the least protected robot is the smallest, and the most protected is the largest, with nothing in between. The cleanroom versions also arrived late. The standard 11 and 15 kg models were announced in June 2022 and available from the end of that year; the 6 and 8 kg builds joined during 2023. The first cleanroom variant, the 11 R1300 CR, appears in the manufacturer's own product listing during 2024, and the 15 R930 CR only in 2025 — three years after the arm it is based on. KUKA publishes no announcement date for either, so those years are lower bounds rather than launch dates. One detail is worth keeping in the manufacturer's favour. KUKA quotes the ISO Class 3 rating without a speed or payload condition, because the CR is a separate catalogue number rather than a standard robot asked to slow down. That is the more expensive way to sell a cleanroom cobot and the one that survives reading the small print. All seven builds are now in the wujec.ai catalogue.
LBR iisy 15 R930 CR →Techman still sells the cobots its own 2022 models beat on every number
Read Techman Robot's six-model TM AI Cobot line as one table and an awkward pattern appears: the two arms the Taiwanese manufacturer released in 2022 outclass the two it released in 2018 in every published parameter, and all four remain in the catalogue. The TM20 carries 20 kg over a reach of 1300 mm and weighs 32 kg. The TM12, four years older, carries 12 kg over exactly the same 1300 mm reach and weighs 33 kg. Same working envelope, eight kilograms less payload, one kilogram more mass on the mounting plate. The other pair repeats the shape: the TM16 carries 16 kg to 917 mm at 32 kg, against the TM14's 14 kg to 1100 mm at 33 kg — here the older arm at least keeps a reach advantage, which is the only column in the whole line where 2018 wins. The naming does not help a buyer see this. In the heavier models the number is the payload, so TM12 through TM20 read as a ladder. In the TM5 pair it is not: the TM5-700 carries 6 kg and the TM5-900 carries 4 kg, both under a name suggesting five, and neither reaches the millimetre figure in its own model number — the datasheets say 746 mm and 946 mm. Techman publishes no list prices, so the catalogue offers no visible reason to choose the 2018 pair over the 2022 pair. Whatever that reason is — price, installed base, spare-part continuity for existing cells — it does not appear on the product pages, and a buyer sizing a workstation from reach and payload alone would have no way to find it. The company's newest line moves in a different direction again. The S series, launched in 2023 and extended in 2025 with the TM6S and TM20S, is sealed to IP65 rather than IP54 and quotes repeatability of plus/minus 0.03 mm, which Techman describes as up to 70 percent better than the earlier arms. Sealing and precision usually trade against each other; here they improved together, one generation apart from the arms that are still on sale beside them.
Techman TM5-900 →The same robot, three different accuracies: how Doosan's spec sheet moved under a fixed model name
A catalogue entry is a snapshot of what a manufacturer says today. Set the same page side by side with its own older versions and the snapshot turns into a record of how a company's confidence in its product changed. Doosan Robotics launched with four collaborative arms - M0609, M0617, M1013 and M1509 - shown together at Robot World in Seoul at the end of September 2017. All four are still on sale under exactly the same names. Their declared repeatability is not the same as it was. In October 2018 Doosan's own product cards gave every one of the four an identical figure: plus/minus 0.1 mm, whatever the arm's length. By 2020 the cards had split: 0.05 mm for the 900 mm M0609 and M1509 and for the 1300 mm M1013, 0.1 mm still for the 1700 mm M0617. Today the pages read 0.03 mm for both 900 mm models, 0.05 mm for the M1013 and 0.1 mm for the M0617. Read as a series, that is a threefold tightening on the short arms across eight years, and no change at all on the long one. It is also the origin of a pattern this catalogue described earlier: across Doosan's current twelve cobots, declared repeatability tracks reach and ignores payload. That rule was not there at the beginning. In 2017 the company published one number for the whole family; the fan-out by arm length appeared later, as the declarations were revised model by model. What the manufacturer does not do is explain why. The product pages carry no revision history, no note of a hardware change and no reference to a measurement standard. Three explanations fit the evidence equally well: the arms were physically improved, the test method changed, or Doosan simply grew confident enough in a mature product to publish a tighter figure. Nothing on the page distinguishes them. The practical lesson is for anyone comparing robots on paper. A specification sheet is dated even when it does not carry a date, and a model name is not a guarantee that the machine behind it - or the claim in front of it - has stayed put. This catalogue records the current figure as the specification and the older ones as history, and says which is which. The four M-SERIES arms now have profiles in wujec.ai.
M0609 →Read all twelve Doosan cobots at once and one rule appears: precision is set by arm length, not by payload
Doosan Robotics sells twelve collaborative arms, and its own line-up page lists each of them with exactly three numbers: payload, reach and repeatability. Put side by side, those thirty-six numbers say something the product pages never state — the third number depends only on the second. Every Doosan arm with a 900 mm reach is quoted at 0.03 mm repeatability. That holds for the A0509 lifting 5 kg, for the force-sensing A0509S, and equally for the M1509, which lifts fifteen kilograms — three times the load, same reach, same precision. The arms that reach 1,200 or 1,300 mm are all quoted at 0.05 mm: the 9 kg A0912, its S and F variants, and the 10 kg M1013. And every long arm is quoted at 0.1 mm: the 20 kg H2017 at 1,700 mm, the 25 kg H2515 at 1,500 mm, and the new P3020, which carries 30 kg to 2,030 mm. The intuitive expectation is the opposite. A buyer asked which of two cobots will be more accurate usually reaches for the lighter one, on the assumption that heavy payloads bend things. Doosan's own catalogue says the load is not what costs you accuracy — length is. A small angular error at the shoulder becomes a large error at the tool, and it grows with the lever arm. Payload is handled by stiffer joints and stronger motors; leverage cannot be engineered away. There is exactly one exception in the line-up, and it is instructive. The E0509 also reaches 900 mm, but repeats to 0.05 mm rather than 0.03 mm. It is the sealed food-industry model, rated IP66 so it can be washed down without a protective jacket — and, as this catalogue noted earlier, sealing the joints is what costs it those two hundredths of a millimetre. It breaks the rule for a reason that has nothing to do with geometry. Seen through this lens, the P3020 that Doosan unveiled at Automate 2024 is not a machine that compromised on precision. It is the least precise arm the company sells, at 0.1 mm, because it is the longest — and it is aimed at stacking cartons on pallets, where a tenth of a millimetre is far below anything the job can notice. The same page shows why the shortest arms are the precise ones and why no amount of paying more will buy 0.03 mm at two metres of reach. None of this is a criticism of the manufacturer's disclosure: the numbers are published openly and consistently, which is what makes the pattern readable at all. It is a reminder that a specification table read one row at a time answers a different question than the same table read as a whole.
P3020 →Doosan sells two cobots for food work. The one you may hose down is the less precise one
Doosan Robotics answers the question of how to put a robot arm near food twice, and its own product pages show what each answer costs. The Korean maker's E-SERIES exists for the food and beverage industry. Its single model, the E0509, carries 5 kg over a 900 mm reach, holds NSF certification for food hygiene and is rated IP66, with the gaps between its axes sealed. Doosan's selling point is what that removes: hygiene control without extra equipment such as robot jackets — the washable covers normally wrapped around an industrial arm working near food, which cost money, hide the joints and get in the way of cleaning. The same catalogue also lists the A0509F, the food-and-beverage build of the lightweight A-SERIES. It has the same 5 kg payload, the same 900 mm reach and the same NSF certification. The difference is in the third number. The A0509F repeats to 0.03 mm. The sealed E0509 repeats to 0.05 mm. Two hundredths of a millimetre is what closing the joints costs, and Doosan publishes no IP figure at all on the A0509F page. That is not a defect, it is a choice presented to the buyer, and an unusually honest one. A plant that cleans by wiping can have the more accurate arm. A kitchen that cleans with a hose has to give up precision to get a machine that survives the water. What the marketing does not say — and what only appears when the two cards are read side by side — is that hygiene ratings and accuracy pull against each other, and the robot advertised as the food specialist is the coarser of the two. The E-SERIES was announced on 17 April 2023, part of a launch of thirteen cobots across five families, and shipped from June that year. Doosan's own listed uses are consumer-facing rather than industrial: loading a fryer and cooking chicken at 170 degrees, pouring hand-drip coffee at three cups in fifteen seconds, cooking noodles at constant quality. Both robots now have profiles in this catalogue. Neither figure for arm mass, joint speed or list price appears on the manufacturer's pages, so those fields stay empty here rather than being filled from distributor listings.
E0509 →Three cobots, three cleanroom ratings — and two of them come with an asterisk
Reading three collaborative arms' data sheets side by side turns up the same sentence written three different ways, and the differences decide whether the number on the brochure survives contact with a real production line. ABB launched its PoWa cobot family from Zurich on 22 April 2026, and the flagship PoWa 13 is the fastest thing in the class: 5.8 m/s at the tool with the full rated payload on the flange, 8.3 m/s with a kilogram, and pose repeatability of 0.03 mm while doing it. The company also certifies the arm for ISO 14644-1 Class 4 cleanrooms. The condition sits one line further down the data sheet: Class 4 applies at 50 percent of rated speed. At full speed the arm is Class 5. The same construction appears at Universal Robots. The UR15, shown at Automate in Detroit in May 2025 and the fastest arm the Danish company has built at 5 m/s, is likewise rated for Class 4 — with a footnote tying the rating to velocity and payload held at or below 80 percent. This catalogue noted that clause when the UR15 profile went up. ABB's derating is the steeper of the two, which is the arithmetic consequence of the faster arm: more speed means more particles shed, so more speed has to be given back. KUKA answers the question differently. Rather than derate the standard machine, the German company builds separate cleanroom variants: the LBR iisy catalogue lists an 11 R1300 CR and a 15 R930 CR next to the standard models, and rates the cleanroom build up to ISO Class 3 — a cleaner class than either rival claims — with no published speed condition attached. That is the more expensive answer for the buyer, because it means purchasing a different robot rather than turning down the one already installed, and it is the one that holds up when the line runs at rate. The practical reading is simple. A cleanroom class printed on a cobot brochure is not a property of the robot; it is a property of the robot at a stated speed. Two of these three tell you that in a footnote, and the third tells you by selling you a second part number. All three arms are now in this catalogue: ABB PoWa 13, KUKA LBR iisy 15 R930 and Universal Robots UR15. Figures are taken from each manufacturer's own data sheet or product page — ABB's PoWa CRB 1910 sheet of 27 May 2026, Universal Robots' technical specification of April 2026, and KUKA's LBR iisy product pages.
PoWa 13 →FANUC's smallest collaborative arm is also its fastest and most precise
In a robot range, the small model is usually the compromise: less reach, less payload, and worse numbers everywhere else as well. FANUC's newest collaborative arm breaks that rule, and the break is visible only when its data sheet is read next to its nearest sibling's. The CRX-3iA went into mass production in January 2026 and reached the North American market on 20 April 2026. It lifts 3 kg over a 692 mm reach and weighs 11 kg. The next size up, the CRX-5iA, lifts 5 kg over 994 mm and weighs 25 kg. Read the rest of both cards and the ranking inverts. Repeatability is ±0.02 mm on the small arm against ±0.03 mm on the larger one. Speed goes the same way: 1,200 mm/s in collaborative mode and 2,500 mm/s in high-speed mode, against 1,000 and 2,000 mm/s. The smallest CRX is the tightest and the quickest in the family — it simply cannot carry much. The mass is the design decision the rest follows from. Eleven kilograms is light enough for one person to carry, so FANUC removed the anchor bolts: the robot stands on a magnetic base, mounts floor, wall, ceiling or at an angle, and detects its own installation angle rather than being configured for it. After a move, the operator touches a wall with the welding wire and the arm works out where it now stands, so the taught program survives the relocation — the step that usually makes moving a cobot expensive. The 3 kg is not an arbitrary floor either. It is enough to hold a welding torch and a seam sensor at once, which is what lets the robot locate the joint and generate its own path instead of replaying a fixed trajectory. FANUC names shipbuilding among the applications: structures too large to bring to a robot cell, where the machine has to come to the work. The arm is the first FANUC robot in this catalogue. FANUC does not publish a list price; the robot is quoted per configuration.
CRX-3iA →One adaptive robot family, two very different machines
Yaskawa sells the Motoman NEXT line as a single idea in four sizes: an industrial arm that recognises what is in front of it, decides and checks the result, instead of replaying a taught path. Read side by side, the four data sheets — edition 06/2026 — describe two different machines wearing one badge. The small pair is light and quick. The NEX7 carries 7 kg, weighs 48 kg and repeats a position to 0.01 mm; its base swings at 375°/s and its tool flange turns at 1,000°/s. The NEX10 carries 10 kg for 58 kg of own mass and 0.015 mm. The large pair is a different kind of hardware. The NEX20 doubles the payload of the NEX10 but nearly quadruples the mass, from 58 kg to 225 kg. The NEX35 reaches 35 kg of payload and 370 kg of mass. The ratio of robot weight to payload jumps from roughly six-to-one in the small sizes to eleven-to-one in the large ones — the series is a shared control philosophy, not one mechanism scaled up and down. Speed and precision follow the mass. From NEX7 to NEX35 the base slows from 375°/s to 180°/s, the tool flange from 1,000°/s to 400°/s, and repeatability loosens fourfold, from 0.01 mm to 0.04 mm. The intelligence is identical across the range — the same YNX1000 controller and the same NVIDIA Jetson Orin decision unit — but a robot that decides in real time and then swings its base at 180°/s is a different proposition from one that does it at 375°/s. Adaptivity is sold uniformly; the ability to act on it is not. One number is not monotonic: the NEX20's tool flange turns at 900°/s, faster than the NEX10's 800°/s, so wrist speed is a design choice per size rather than a straight consequence of payload. Two caveats about the source. Yaskawa's own NEX35 sheet is inconsistent about the product name — the headline reads NEXT35, while the file name, the running head and the catalogue all read NEX35 — and it prints the lower-arm motion range as -140/-90°, a span of 50° that no articulated lower arm has; neighbouring sizes are given as +135/-100° and +145/-100°. We have treated that figure as a misprint and left it out of our profile rather than repeat it. All four sizes now have profiles in our catalogue. The gap between what was announced and what is offered runs wider than one model: the 4 kg NEX4 of the 2023 announcement has no product page in Yaskawa's North American catalogue, and neither do the two collaborative arms, NHC12 and NHC30, presented with the platform in September 2024 — the collaborative section of that catalogue lists only the older HC10DTP, HC20DTP and HC30PL (checked 6 September 2026). Seven machines were announced for the platform; four can be found in the catalogue.
Motoman NEXT NEX35 →Yaskawa promised five adaptive robots. The smallest one has no page to buy it from.
When Yaskawa announced the Motoman NEXT series in November 2023, it called it the industry's first industrial robot line able to judge a situation instead of replaying a taught path, and it named the lineup precisely: five models with payloads of 4, 7, 10, 20 and 35 kg, on sale from December 2023. The American launch in September 2024 repeated the same five and added two collaborative arms, NHC12 and NHC30. The catalogue looks different today. Yaskawa America's NEX series listing, checked on 6 September 2026, shows four robots: NEX7, NEX10, NEX20 and NEX35. There is no NEX4 tile, and the product address the series URL pattern would give it returns a 404 error page rather than a data sheet. The gap matters more than one missing tile. Adaptive behaviour — vision, force sensing, deciding what to do with a part that is not where it should be — is worth most on small, fiddly work: electronics assembly, laboratory handling, packing awkward little objects. That is the 4 kg class. A lineup that starts at 7 kg starts above the applications the technology was sold on. What the absence means is not stated anywhere. Yaskawa has published no withdrawal notice for the NEX4, and the model may still exist in other regions or as a special order; a product page can also vanish in a website rebuild. Until the company says otherwise, the honest reading is narrow: as of today, a North American buyer who wants the smallest adaptive Motoman has nowhere on Yaskawa's own site to start. Our catalogue profile of the NEX10, the middle size of the same family, records the same finding in its availability field.
Motoman NEXT NEX10 →Two Cyberdyne exoskeletons, two very different meanings of "FDA"
Cyberdyne sells two medical exoskeletons in the United States, both branded Medical HAL, both controlled the same way — and each stands on a different regulatory footing. One went through the FDA's 510(k) review three times. The other is merely registered as a Class I device. On a marketing page both can be described as being with the FDA. The leg device, Medical HAL Lower Limb Type, holds three 510(k) clearances in the agency's own database: K171909 on 17 December 2017, K201559 on 2 October 2020 and K233695 on 7 May 2024. That route requires the manufacturer to show the agency that the device is substantially equivalent to one already on the market, and it is repeated whenever the product changes materially — which is why there are three entries rather than one. The single-joint device, HAL-SJ, takes the other path. Cyberdyne states on its own product page that the unit is CE marked under the older European Medical Device Directive and registered with the FDA as a Class I device. Class I registration is a listing obligation, not a review of the device; for most Class I products no clearance decision is made at all, and none appears in the 510(k) database. Neither position is irregular — Class I is the correct category for a large family of low-risk devices, and the two HAL products are not equivalent machines. The 1.5 kg single-joint brace drives one elbow, knee, ankle or shoulder at a patient's bedside; the leg exoskeleton weighs up to 14 kg and is used for gait training in spinal cord injury, stroke and cerebral palsy. What the pairing shows is how little the phrase "FDA" carries on its own when it appears in a robotics spec sheet. Two devices from one manufacturer, sharing a control principle and a product name, sit on opposite sides of a line the buyer cannot see from the brochure. The catalogue records the specific basis for each: three clearance numbers with dates for the leg device, Class I registration for the single-joint one.
Medical HAL Single Joint Type →Japan's companion robot gets 12% dearer — and the sticker was never the price
GROOVE X announced on 1 September 2026 that LOVOT 3.0, its wheeled companion robot, will cost more in two steps: the base colour goes from JPY 577,500 to JPY 599,500 on 26 October 2026, and to a planned JPY 649,000 in mid-January 2027. The company gives the reason plainly — rising prices of raw materials and parts, and higher manufacturing and logistics costs — and says the four-month, two-stage schedule is meant to give buyers time to decide. The older LOVOT 2.0 stays at JPY 449,900. That is a 12.4 percent rise on a robot that does nothing useful by design: it does not clean, carry or answer questions, it asks to be picked up. What the announcement does not say is that the purchase price has never been the whole cost. LOVOT cannot be used without a monthly plan, which carries the software, the cloud services and the servicing: JPY 9,900 a month at the cheapest tier, JPY 12,980 for the tier GROOVE X itself recommends, JPY 19,800 for full cover. Put the two together and the proportions change. At the recommended tier the subscription costs JPY 155,760 a year. Four years of it — roughly the interval at which the price list schedules a servo replacement pack — comes to JPY 623,040, more than the new purchase price of the robot itself. The announced increase raises the entry ticket by JPY 71,500; a single year of the plan costs more than twice that. None of this is hidden: the figures come from the company's own price list, which also states the servicing intervals and the colour surcharges of up to JPY 122,100. It is worth stating clearly all the same, because a headline price is what gets compared. Against a Chinese humanoid that keeps getting cheaper, LOVOT looks expensive. Against the four-year cost of keeping one alive, the sticker is the smaller half.
LOVOT 3.0 →Meituan flies 20 km with a winch at home and exports half the machine
Meituan runs the largest food-delivery drone fleet in the world, and it sells abroad under a separate brand: Keeta Drone, which operates in Dubai and Hong Kong. Its technology page lists exactly one aircraft, the Keeta Drone Gen 4: 7.2 kg empty, 2.4 kg of payload, 10 km of range, 10 m/s cruise. That is the standard model. In China the same fourth generation has a second member, the M-Drone 4L Winch, and the gap between them is not a detail. The winch version carries 4.5 kg instead of 2.4, flies up to 20 km fully loaded instead of 10, and lowers the parcel on a cable from a hover, so it does not need a prepared landing cabinet at the other end. Its stated weather envelope is heavy rain, moderate snow and Beaufort 6 wind, down to -20 °C. None of that is on offer to an export customer today. A partner signing up in the Gulf gets an aircraft with roughly half the payload and half the range of the strongest machine in the manufacturer's own fourth generation, and one that still depends on a fixed pick-up cabinet at the delivery end. There are unglamorous reasons why this can be so. A winch that pays out a cable over a street is a harder case for a regulator than a drone that lands inside a locked box on a rooftop, and export approvals are usually granted for one configuration at a time. Cold-weather figures also matter little in a country where the problem is 50 °C, not -20 °C. Still, the practical conclusion is worth stating plainly: the numbers quoted in coverage of Meituan's Chinese operations do not describe the aircraft on sale outside China. Anyone comparing drone delivery services should check which of the two is actually flying overhead.
Meituan M-Drone 4L Winch →The label says Apache. One file out of four says no.
ThinkSound, the video-to-audio model from Alibaba's audio lab, is listed on Hugging Face under Apache 2.0 — the licence developers treat as a green light: use it, ship it, sell what you build with it. The repository even contains the full Apache text. Anyone who stops reading there will get it wrong twice over. The first correction is at the bottom of that same file, and repeated in the project README: the code, models and dataset are for research and education only, and commercial use is not permitted. A restriction of that kind sits oddly inside a document whose whole point is to remove restrictions, but it is the authors' stated intent, and it is stated twice. The second correction is more interesting, because it would survive even if the authors changed their minds. The release is four files: a 21.06 GB generator, a 5.73 GB lightweight version of it, a 950 MB synchronisation module — and a 2.52 GB audio autoencoder. That last one is not Alibaba's work. It is a fine-tuned copy of Stable Audio Open, released by Stability AI under the Stability AI Community License, which explicitly forbids redistribution under a different licence and reserves commercial use above a revenue threshold. The notice in the repository says so plainly. What makes this more than a footnote is what the file does. The autoencoder is the part that turns the model's internal representation into an actual waveform. Without it the other three files produce nothing you can listen to. So the most restrictive licence in the package is attached to the component that cannot be removed — and it governs the working system, not an optional extra. The pattern is not unique to this release. PrismAudio, the successor from the same team and the same repository, carries an MIT label and a model card restricting the weights to research and education. In both cases the machine-readable field says one thing and the human-readable text says the opposite, and it is the field, not the text, that search filters and automated compliance checks read. For a studio wondering whether it can use these models on paid work, the answer is no in both cases. For everyone else, the useful habit is smaller and more general: on a model page, the licence tag is a claim, not a document. Open the file it points to and read to the end — and check whose weights are in the box alongside.
ThinkSound →In voice models, the number in the name counts only the part that thinks
A text model called 8B holds about eight billion parameters. A voice model called 7B can hold ten and a half billion. We checked the published weight files of nine models and the gap turns out to be the rule, not an exception: the figure in the name describes the language backbone, while the encoder that hears and the head that speaks are simply not counted. The check is mechanical. Every model published on Hugging Face reports the exact size of its weight files, and the configuration file next to them shows what those weights are made of. Against a text control — Meta's Llama-3.1-8B-Instruct, which holds 8.03 billion parameters and is accurate to within half a percent — the voice models line up like this. Alibaba's Qwen2.5-Omni-7B holds 10.73 billion, fifty-three percent more than its name. Moonshot's Kimi-Audio-7B-Instruct holds 9.77 billion, forty percent more. LLaMA-Omni2-7B holds 8.95 billion. Qwen3-Omni-30B-A3B, downloaded more than eight hundred thousand times in the past month, holds 35.26 billion. Fun-Audio-Chat-8B, the model Alibaba opened in December and which we added to the catalogue today, holds 9.45 billion. What gets left out is always the same two organs. Fun-Audio-Chat's configuration file shows a text backbone with exactly the geometry of Qwen3-8B — thirty-six layers, 4096 hidden size, a vocabulary of 151936 tokens — accounting for 8.19 billion parameters. Above it sits an audio encoder of thirty-two layers, 1280 wide, twenty attention heads, reading 128 mel bands in thirty-second windows. That is, parameter for parameter, the shape of the Whisper-large encoder, and it reappears across vendors: the Ultravox project loads a component of identical geometry by name. Beside the backbone sits a third module, a speech head with the dimensions of Qwen3-0.6B, whose job is to emit sound. Roughly 1.25 billion parameters of ears and mouth, absent from the label. For a reader choosing hardware this is not trivia. In bfloat16, Llama-3.1-8B-Instruct occupies 15.0 GiB and fits on a 16 GB card. Qwen2.5-Omni-7B, whose name promises something smaller, occupies 20.0 GiB and does not. Qwen3-Omni-30B-A3B needs 65.7 GiB before a single second of audio is loaded. Fun-Audio-Chat's producer states the practical consequence plainly in its requirements: about 24 GB of GPU memory, and a second checkpoint downloaded separately if you want the model to speak rather than write. The naming is not dishonest everywhere. Microsoft's Phi-4-multimodal-instruct advertises 5.6 billion parameters and delivers 5.57 billion, and its card states in one sentence that the figure covers the Phi-4-Mini backbone plus the vision and speech encoders and adapters. It can be done. The error also runs the other way, which is the reason to check rather than to assume a direction. The repository fixie-ai/ultravox-v0_5-llama-3_1-8b has an eight in its name and contains 0.69 billion parameters — 1.3 GiB. The Llama backbone the name refers to is not inside; it is pulled separately at load time, and what ships is the hearing apparatus and the adapter. Downloading the file named after an 8B model gets you, in that case, everything except the 8B model.
Fun-Audio-Chat-8B →The most downloaded speech model at Alibaba's audio lab is also the only one you can lose for criticising it
FunAudioLLM, the audio laboratory inside Alibaba's Tongyi group, publishes twenty-one models on Hugging Face. Twenty of them carry Apache 2.0. The exception is SenseVoiceSmall — and it is the one the public actually uses, with 32,530 downloads in the past thirty days against 4,253 for the second-place recogniser. Its licence label reads simply "other", and the link behind it leads to a four-page document that most users of an open model would not expect to have agreed to. The document is the FunASR Model Open Source License Agreement, version 1.1, and three of its clauses deserve attention. Clause 4.2, headed Prohibited Behavior, states that users "shall not engage in unjustified denigration, malicious smearing, or baseless insults" against the software, and that doing so "will be considered an automatic forfeiture of all licenses under this agreement". Clause 5 confirms the mechanism: a breach terminates the licence automatically and the user must stop using, copying, modifying and sharing the model. A licence to run a speech recogniser is therefore conditional, in writing, on how its holder speaks about that recogniser in public. Clause 3 pulls in a different direction from the marketing. The model card presents SenseVoice as part of an "industrial-grade" toolkit; the licence says the software "is provided for reference and learning purposes only". Whether that phrase is a disclaimer of liability or a restriction on the field of use is not resolved anywhere in the text, and the question is not academic for anyone deploying the model in a product. Clause 7 leaves the reader with nothing to resolve it against. It reads: "This agreement is governed by the laws of [Country/Region]." The square brackets are in the published file. This is not a translation artefact — the Chinese half of the same document carries the identical unfilled placeholder, 国家/地区. Clause 6 adds that Alibaba may revise the agreement at any time, that the revision takes effect automatically on publication in the repository, and that continued use signifies acceptance. None of this makes the model unusable, and the practical restriction — attribution, retention of model names — is mild. But it is a materially different bargain from the Apache 2.0 under which the same laboratory ships its newer work, including the December 2025 Fun-ASR-Nano pair and the Fun-CosyVoice 3.0 speech engine. A team that downloads one FunAudioLLM model after another, seeing Apache on each, has no cue that the most popular file in the set arrives on other terms.
SenseVoiceSmall →The new open transcriber beats Whisper — until you leave the big languages
Alibaba's Qwen3-ASR is being read as the model that finally unseats Whisper-large-v3, the OpenAI transcriber that has been the open default since 2023. On the headline numbers it does: 4.90 per cent word error rate against 5.27 on the Fleurs benchmark. But the producer's own results table contains three lines called Fleurs, Fleurs† and Fleurs††, and they are not three tests of the same thing — they are the same test run over 12, 20 and 30 languages. Which line you quote decides who wins. Over the twelve largest languages — English, Chinese, Cantonese, Arabic, German, Spanish, French, Italian, Japanese, Korean, Portuguese, Russian — the new model wins, 4.90 to 5.27. Add eight more, including Polish, Dutch, Hindi, Turkish and Vietnamese, and the win shrinks to a statistical draw: 6.62 to 6.85. Add the last ten — Czech, Danish, Greek, Persian, Finnish, Filipino, Hungarian, Macedonian, Romanian, Swedish — and the order reverses hard: 12.60 against Whisper's 8.16. The averages let the last group be estimated. If each language weighs the same, the ten smallest work out to roughly 25 per cent errors for the new model against about 11 for the three-year-old one — one word in four against one in nine. That is the difference between a transcript a person edits and a transcript a person retypes. None of this makes the release weaker than it looks in the other direction: the same model handles streaming and offline transcription, covers 22 Chinese dialects, transcribes singing and songs over backing music, identifies the spoken language more reliably than Whisper (97.9 against 94.1 per cent), and ships under Apache 2.0. The point is narrower and it is the one a reader actually needs: a state-of-the-art claim in speech recognition is only as broad as the language list underneath it, and that list is usually a footnote. Anyone working in Czech, Hungarian, Romanian or the Nordic languages should test the old model before replacing it. Method note: the tier figures are the producer's own; the estimate for the last ten languages is a wujec.ai calculation from the published averages and assumes each language is weighted equally.
Qwen3-ASR →Alibaba publishes four small models and quietly stops recommending two of them for real work
Vendors rarely tell you where their own product stops being useful. Alibaba does, in one sentence, and it is easy to miss. The Qwen3.5 generation runs from a 2.4-trillion-parameter flagship down to a model of 800 million. We have now catalogued the bottom three rungs — Qwen3.5-4B, Qwen3.5-2B and Qwen3.5-0.8B — and reading their model cards side by side shows a line the company draws through its own line-up. At 4B and above, the card describes a product. At 2B and below, it adds a sentence that appears nowhere higher up: given the parameter scale, the intended uses are prototyping, task-specific fine-tuning, and research or development. In the same place, a second thing disappears. The 27B, 9B and 4B cards all promise a context window of 262,144 tokens extensible to 1,010,000. The 2B and 0.8B cards promise the 262,144 and say nothing about the million. The benchmark tables show why. On text, the 0.8B scores 11.9 on GPQA and 8.2 on PolyMATH, and on the HMMT competition-mathematics sets Alibaba reports no figure at all — the dashes in its own table are the honest answer. Instruction following lands at 44.0, meaning a complicated prompt is not reliably obeyed. What does not collapse is perception. Every model in this generation is a vision-language model, and the small ones read remarkably well for their size. The 2B posts 84.5 on OCRBench, beating not only the previous generation's dedicated 2B vision model (79.2) but its 4B one as well (80.8). The 0.8B, at roughly 1.7 GB of weights, reaches 74.5 — and 79.1 with thinking switched off, which is level with a dedicated vision model more than twice its size. That is the practical reading. Below two billion parameters, Alibaba's own documentation says: use these to read, label and extract, or to fine-tune for one narrow job. Do not use them to think. It is an unusually candid piece of labelling in a field where every release is described as a breakthrough, and it is worth more to a reader choosing a local model than any leaderboard position.
Qwen3.5-2B →The slope figure on a robot mower's spec sheet often stops at the edge of the lawn
Robot mowers are sold on a single slope number, and buyers with a sloping garden read it as the machine's limit. In the manuals it is frequently two numbers, and the second one — the slope the mower may take at the boundary of the lawn — can be a third of the advertised figure. Segway-Ninebot's new Navimow H2 series, which reached the manufacturer's European store on 9 February 2026, is rated for 45 per cent inside the working area and 25 per cent at the boundary. The i2 LiDAR declares exactly the same split. Husqvarna is blunter still: the Automower 450X is specified at 45 per cent inside the installation and 15 per cent at the boundary, a threefold difference. Within Segway's own catalogue the split follows the drivetrain rather than the price. The i2 AWD, the cheapest machine in the line at EUR 899, keeps its full 45 per cent right up to the edge. So does the i2 LiDAR Pro at 55 per cent, and so does the flagship X4, which holds 84 per cent — 40 degrees — both inside the lawn and at its boundary. The two models that lose ground at the edge, the H2 and the i2 LiDAR, are the two without four-wheel drive. A better sensor package does not buy traction on a bank. The distinction matters because the boundary is where sloping gardens usually slope. A lawn that falls away towards a fence or a ditch meets its steepest gradient at precisely the point the mower is rated for least, and an owner who checked only the headline number will have bought for the wrong part of the garden. The figures above are all taken from the manufacturers' own specification tables — Segway publishes them only inside the user manuals, not on the product pages.
Segway Navimow H2 Series →One Segway mower series, three ways of knowing where it is — and only the cheapest one may end up with an antenna in your lawn
Segway-Ninebot now sells three robotic lawn mowers under one name, the Navimow i2. They look like a price ladder: i2 AWD from EUR 899, i2 LiDAR from EUR 1,199, i2 LiDAR Pro from EUR 1,599. Read the three user manuals side by side and they turn out to be three different answers to the same question — how does a mower know where it is in a garden? The cheapest machine has no lidar. The i2 AWD finds itself with satellite corrections delivered over its own cellular link, which Segway calls network RTK, fused with a forward camera that fills in the gaps where the signal fades. The middle model drops satellites and maps the lawn with a laser scanner and a camera instead. The top model, sold since 16 March 2026, is advertised as fusing all three: lidar, network RTK and vision. The consequence for a buyer is in the installation chapter, and it runs the opposite way to the price. The i2 AWD manual is the only one of the three that contains a section headed Install the Antenna. Where the network correction service is not supported in a given area, the owner has to plant the supplied reference antenna in an open spot with an unobstructed view of the sky and at least two metres clear on either side, or buy a mounting kit and fix it to a wall. The two lidar manuals have no such chapter, no antenna and — checked word by word — not a single mention of RTK anywhere in their pages, even in the model Segway markets as using it. Their installation chapters amount to switching the machine on, assembling the charging station and updating the software. So the marketing line about an antenna-free lawn is true of the whole series, but it is unconditional only on the expensive models. On the cheap one it holds inside the coverage map and not outside it. Buyers on the edge of network coverage are exactly the buyers most likely to be shopping at the bottom of the range. One further oddity emerged from the same manual. The store lists two i2 AWD variants, the i205 for 500 square metres and the i208 for 800. The manual specifies four, adding an i206 for 600 square metres and an i210 for 1,000 — and the store does sell extended warranty plans for both machines it does not otherwise admit to having. All three mowers are now in the catalogue, with the specifications taken from the manufacturer's own manuals.
Segway Navimow i2 AWD →Segway sold four mowers as one family — only the top one could see an obstacle before hitting it
Segway's H Series is presented as one machine in four sizes: the same body, the same 21 cm cutting width, the same 30 to 60 mm cutting height, the same 54 dB(A) sound power. The only thing the marketing separates is lawn area — 500, 800, 1,500 or 3,000 square metres — and the battery that goes with it. The manufacturer's own manual separates two more things, and neither of them appears in the sales copy. The first is the ultrasonic sensor. In the packing list of the H-series manual it carries a footnote of its own: "For H3000E". The specification table backs this up with two different sensor columns — three of the four machines get an IMU, a blade-halt sensor, a bump sensor, a lift sensor, a rain sensor and wheel encoders, and only the largest model adds ultrasound. For the buyer this is the difference between a mower that notices an object at a distance and a mower that finds out about it by driving into it and reversing. The second is the cellular modem. The status display chapter states that the 4G indicator applies "for H800E, H1500E and H3000E models", which leaves the entry-level H500E on Wi-Fi and Bluetooth alone — the one machine most likely to be bought for a small garden at the far end of a house's wireless coverage. Neither gap is hidden, exactly. Both are simply printed in a 69-page PDF rather than on the product page, and both survived the line's 2025 relaunch as the camera-equipped H500-VF through H3000-VF, priced from 1,399 to 2,699 euros. As of September 2026 every H variant shows as unavailable in Segway's European store, which now sells only the lidar-navigated i2 range and the X4 flagship. The pattern is worth carrying forward: in a family sold on lawn size, the cheaper number can also buy a shorter sensor list.
Segway Navimow H Series →The same Mammotion mower ships with an RTK station in Europe and is advertised in America as needing none
Mammotion's YUKA mini is one product name covering two different machines, and the difference is the part a buyer cares about most: how the mower knows where it is. The European store sells the 600 and the 800, and the packing list on the product page is explicit. In the box, next to the mower, its charging station and a spare set of blades, there is an RTK reference station, an RTK antenna, a power supply for it, a five-metre extension cable and a ground pile to drive into the lawn. Satellite corrections come from hardware the owner installs and points at the sky. The American store sells the same mower as the 600H and the 800H, and its listing opens with the opposite sentence: no wire, no RTK installation. There, corrections arrive over the network through Mammotion's iNavi service, which the company describes as centimetre-level positioning without a physical base station. The footnote underneath matters: the American version has no built-in 4G module, so the network link that the whole arrangement depends on is a paid bundle, not part of the machine. The cameras are identical on both sides. Mammotion calls the system UltraSense AI Vision, quotes an AI chip running five trillion operations per second, and claims the mower can keep cutting for up to three hundred metres with the satellite link disrupted, syncing back up when it returns. That claim reads differently depending on which box arrived: in Europe it is a fallback for when the garden's own base station loses sight of satellites, in America it is a fallback for patchy mobile coverage. The practical consequence is for anyone shopping by review. A European owner writing about installing an antenna on a pole and an American owner writing that the mower needed no installation at all are both describing the YUKA mini accurately. Prices differ too: EUR 599 and EUR 699 in Europe against a EUR 899 list, with both variants sold out there by September 2026, and USD 749 and USD 1,299 in the American store, which still has stock. As with every cheaper mower from this manufacturer, the published figures stop early. Mass is given as 10.6 kilograms in Europe and 23 pounds in America, slopes at 50 percent, cutting height between two and three and a half inches. Cutting width, dimensions, mowing time per charge and noise appear nowhere.
Mammotion YUKA mini →In both of Mammotion's compact mower lines, the cheaper variant is not a smaller machine — it is the one that swapped lidar for a mobile signal
Mammotion's 2026 catalogue names its compact mowers after lawn area: LUBA mini 2 AWD 1000 and 1500, YUKA mini 2 800 and 1000. The numbers suggest the only difference is how much grass the machine is rated for. Read the manufacturer's own product pages for all four and a different line appears, and it runs through the sensors. In each line, the more expensive variant carries a 360-degree lidar. The LUBA mini 2 AWD 1500, listed at EUR 1,699, positions itself from lidar and two cameras and needs no network connection of any kind. The YUKA mini 2 1000, at EUR 1,099, is sold under the headline "Ultimate 360 LiDAR" and the flat statement "No RTK, No Cable". In each line, the cheaper variant has no lidar at all. The LUBA mini 2 AWD 1000, at EUR 1,399, fuses three cameras with NetRTK — satellite corrections delivered over 4G or Wi-Fi instead of from an antenna in the garden. The YUKA mini 2 800, at EUR 899, does the same with three cameras and Mammotion's iNavi service, which the company advertises as free for the life of the machine. That is a real difference in what the buyer has to have at home, and the model name does not carry it. A lidar machine measures the garden itself and works the same whether or not the phone network reaches the far hedge. A camera-and-NetRTK machine has one absolute reference, and it arrives over the air. Trees, walls and dusk are handled by the cameras in both cases; a dead spot in mobile coverage is not the same problem for both. The wording on the cheaper pages is worth reading closely for the same reason. The YUKA mini 2 800 page says "No RTK, No Cable, No signal issue" and, a few sections later, offers lifetime free access to a network correction service that works "without the need for a physical base station". Both are true in the narrow sense the manufacturer means — there is no base station to install and no antenna to mount — but a reader who takes "No RTK" at face value would conclude the machine ignores satellites altogether, and it does not. None of this makes the cheaper machines bad buys. They cost EUR 200 to EUR 300 less, they install in minutes, and Mammotion throws in three years of mobile data on the YUKA mini 2 800 precisely because the link matters. It is a straightforward trade, stated plainly here because the product names state it nowhere: in this catalogue the price step is not a step in lawn size, it is the step between a mower that measures the garden and a mower that is told where it is.
Mammotion YUKA mini 2 800 →Mammotion calls three-source navigation its breakthrough — and left one source out of the successor to the mower that introduced it
Mammotion sells its compact mowers on a navigation idea it calls Tri-Fusion: solid-state lidar, network RTK corrections and camera-based AI vision working at the same time, so that a garden with trees, walls or narrow passages does not break positioning. The company is explicit about where the idea started. On the product page of the LUBA mini AWD LiDAR, listed in its European store on 30 June 2025, it writes that the system was first introduced on this machine and now powers the 2026 flagship LUBA 3 AWD. That machine is now sold out in Europe, and the store sends buyers to its direct successor, the LUBA mini 2 AWD 1500, launched on 8 May 2026. That variant navigates by lidar and cameras alone. There is no RTK antenna to mount and no network RTK service behind it — the manufacturer presents this as a feature, because it removes the one installation step that has always made satellite-corrected mowers awkward. Both statements are true at once, and together they are more interesting than either alone. In the space of eleven months the same company introduced a third navigation source in its compact line, moved it up to the flagship, and dropped it from the compact line again. A buyer comparing the two mowers is not choosing between an old model and a better new one: the older machine positions itself from three independent sources, the newer one from two, and only the older one can fall back on satellite corrections when lidar and cameras are both struggling — heavy rain, low sun straight into the lenses, a lawn that looks the same in every direction. The trade is real in the other direction too. Network RTK on the older mower is a paid service and one more thing to configure; the newer one works out of the box. Neither approach is simply better, which is why the difference deserves to be stated rather than buried under a marketing name that suggests the newer machine has more of everything. There is a second twist inside the successor line itself, and it is the kind a buyer only finds after paying. The cheaper variant, the LUBA mini 2 AWD 1000, listed in the same store on 1 January 2026 at EUR 1,399, is not the 1500 with a smaller battery. It keeps network RTK and drops the lidar instead, positioning itself from three cameras plus satellite corrections carried over a 4G or Wi-Fi link. So two mowers sold under one series name navigate in opposite ways: the more expensive one needs no network at all and sees the garden in three dimensions; the cheaper one needs a working connection and sees it only through lenses. The series name says nothing about which of the two is on the shelf. The list prices do not track the sensor count either. The older mower was listed at EUR 2,299 and cleared at EUR 1,399; the newer one sells at EUR 1,699.
Mammotion LUBA mini AWD LiDAR →Three generations of CogVideoX, three degrees of closing: Apache 2.0, a revocable licence, then no weights at all
Z.ai's video line has shipped three generations in two years, and each one has been harder to use freely than the one before. The catalogue now holds all three, so the sequence can be read in one place. On 6 August 2024 the producer published CogVideoX-2B, its first video model with downloadable weights: six seconds of 720x480 at eight frames per second, from 4 GB of video memory in the diffusers library and 3.6 GB with INT8 quantisation. It shipped under the producer's own restrictive document. Three weeks later, on 27 August 2024, one entry in the project's changelog did two things at once. It announced the larger CogVideoX-5B — same six seconds, same 720x480, about 5.6 billion parameters in the video transformer instead of 1.7, from 5 GB of memory in BF16 — and, in its closing line, moved the 2B to Apache 2.0. The bigger model did not get the same treatment. It went out under the CogVideoX Licence, which grants a copyright licence that is expressly revocable, permits academic use freely, and requires commercial users to register with the producer and stay below one million service visits a month. So the small model became genuinely free on the exact day the large one arrived without that freedom. That is the opposite of the usual pattern, where a producer opens the small model as a sample and holds the flagship back — here the flagship was released, but on terms the smaller model had just escaped. The third step removed the choice. CogVideoX-3, launched in July 2025, has no public weights at all. It is a hosted product: up to 3840x2160, five or ten seconds, 30 or 60 frames per second, an optional generated soundtrack, a flat $0.20 per video. Whoever built on the open versions cannot follow the line forward — they can only start again on someone else's weights, or pay per clip. The open two are not abandoned. In the 30 days to 4 September 2026, CogVideoX-2B was downloaded 20,387 times and CogVideoX-5B 17,161 times on Hugging Face — two years after release, and a year after the generation that replaced them stopped being downloadable at all.
CogVideoX-5B →Alibaba sells its open video models as bilingual — on one of them, its own card recommends writing in Chinese
Every model card in Alibaba's open Wan 2.1 video family lists two languages, English and Chinese, and says nothing further about the difference between them. One card breaks the pattern. Beneath the download table of Wan2.1-FLF2V-14B-720P, the model that builds a shot between a fixed opening frame and a fixed closing frame, the maker writes that for first-last frame generation it trained the model primarily on Chinese text-video pairs and therefore recommends using a Chinese prompt for better results. The sentence is a single line in a note that otherwise discusses resolution, and it is the only place in the family's documentation where the two advertised languages are described as unequal. The four other builds of the generation - text-to-video at 1.3 and 14 billion parameters, and image-to-video at 480p and 720p - carry the same bilingual label with no such qualification. It matters because of what the model is for. First-last-frame generation is the closest thing the open field has to a storyboard tool: the user fixes both ends of a shot and the model invents the movement between them, which is also the only way to chain clips without each one drifting somewhere the next cannot pick up. That is the build an English-speaking studio would reach for first, and it is the build whose documentation says the English side of the training data was thinner. The usage figures suggest the model is admired more than it is run. It is downloaded around 1,800 times a month, a fraction of the plain image-to-video builds, while carrying 228 bookmarks - a high ratio of interest to traffic. Whether the language note is part of the reason is not something the download counter can answer. wujec.ai has added profiles for this model and for three other first-generation Wan builds the catalogue was missing: the 720p and 480p image-to-video models, published on 25 February 2025, and the small VACE 1.3B video editor from 13 May 2025. All four are Apache 2.0 by label, though the licence file the cards link to is present in two of the four repositories and absent from the other two.
Wan2.1-FLF2V-14B-720P →Two thirds of Z.ai's model repositories are older than 2025. One of them carries most of what is left of their traffic.
Z.ai keeps 154 model repositories on Hugging Face. Ninety-eight of them - just under two thirds - were published before 2025, and together they were downloaded 810,718 times in the last thirty days. That is eight per cent of the account's 10.08 million downloads. The fourteen repositories created in 2026 took 8.38 million, or eighty-three per cent. The surprise is how concentrated the remainder is. ChatGLM2-6B, published in June 2023, accounts for 435,810 of those downloads on its own - more than half of everything the pre-2025 archive collects, and more than the vendor's current GLM-5, GLM-5.1, GLM-5.3 or GLM-4.7 individually pull in the same period. A three-year-old six-billion-parameter chat model outranks four of the company's own flagships on the download counter. Size is the readable explanation. ChatGLM2-6B runs on a 6 GB graphics card at INT4 quantisation and holds about 8,000 tokens of conversation there; the 2026 flagships do not fit consumer hardware at all. The rest is inertia: tutorials, university course material and fine-tuning recipes written in 2023 still point at that repository by name. One caveat belongs with the numbers. Hugging Face reports downloads over a rolling thirty-day window only, so these figures describe present-day pulls, not lifetime totals - a model released last week and a model released three years ago are measured on the same month. wujec.ai has added profiles for ChatGLM-6B, ChatGLM2-6B and GLM-4-9B-Chat, the three models this history runs through.
ChatGLM2-6B →The step height on the box is measured on a staircase. At a doorway threshold, the robot clears about half of it
Robot vacuums with lifting chassis are sold on one number: how high an obstacle they climb. Two manufacturers' own specification tables show that the advertised figure is measured on a double step — two edges in a row — and that a single threshold, the thing actually found in a doorway, has a separate and much lower limit. Dreame's X50 Ultra, added to our catalogue today, advertises 60 mm. The company's own table qualifies that as a two-step climb, and its footnote adds a condition: the gap between the two steps has to be wider than 40 mm. For a single vertical obstacle Dreame states 42 mm. The 2026 successor, the X60 Ultra, repeats the pattern with a wider spread: 88 mm on a double step, 45 mm on a single one. Ecovacs describes its DEEBOT X12 OmniCyclone the same way, in inches: single thresholds up to 0.94 in (24 mm), continuous double steps up to 1.57 in (40 mm). In all three cases the headline number is between 1.4 and 2 times the single-threshold figure. The distinction is not a trick of wording — a chassis that lifts its front wheels onto a first edge can use that edge to lever the rear wheels onto a second one, which is why a staircase-like pair is easier than one tall lip. It is, however, the wrong number for the buyer's question. Doorway thresholds, balcony rails and shower kerbs are single obstacles. Read against the smaller figure, the segment's leaders clear 24 to 45 mm, not 40 to 88 mm. All figures in this note come from the manufacturers' own specification tables and footnotes, and all are described by them as in-house laboratory results.
Dreame X50 Ultra →Z.ai's two most downloaded models carry an MIT badge and no licence file — and so do 35 more
Z.ai publishes 154 model repositories on Hugging Face. Forty-nine of them carry an MIT licence label. Thirty-seven of those forty-nine contain no licence file at all — not LICENSE, not LICENSE.txt, nothing. The label is a line of metadata typed into the model card header, and behind it there is no licence text to read, no copyright holder named, no year. This is not a long tail of abandoned experiments. The two most downloaded models on the entire account are in the list: GLM-OCR, pulled two million times in thirty days, and GLM-4.7-Flash, pulled 1.88 million times. So are GLM-4.5, GLM-4.6, GLM-4.7, GLM-5 and their quantised builds, the vision line GLM-4.6V, the speech model GLM-ASR-Nano-2512, the image model GLM-Image and December's phone-operating agent AutoGLM-Phone-9B. Counted by traffic rather than by repository, the models with a badge and no file account for 5.51 million of the 9.09 million downloads Z.ai's MIT-labelled weights recorded in the past thirty days — three downloads in five. The company clearly knows how to ship a licence when it wants to. GLM-5.2, GLM-5.1 and GLM-5.3-Flash all carry the MIT text in the repository. The flagship GLM-5.3, released on 25 August, ships a bespoke document — the GLM-5.3 License, MIT-like in substance but with a clause requiring a Z.ai security review from any model-as-a-service operator above ten billion dollars in group revenue. Where the company had something specific to say, it wrote it down. AutoGLM-Phone-9B shows what the gap costs a reader. The weights repository is labelled MIT with no file. The code repository on GitHub is Apache 2.0. And the project's own notice states that the work is intended for research and study only, and prohibits use for unauthorised data access or system interference. Three signals, three different scopes, and nothing that reconciles them. A team that picks the model off the MIT badge is relying on a label the publisher never backed with text. None of this makes the weights unusable, and none of it implies bad faith — an omitted file is most easily explained by an upload script that never copied one. But a licence badge is the single field most readers use to decide whether they may ship something, and on this account it is unsupported three times out of four. Until the files appear, the safe reading of an MIT badge in these repositories is that the terms have not actually been published.
AutoGLM-Phone-9B →Z.ai closed the weights of its own video line — then built its open animation model on a rival's
Z.ai has two video products and they point in opposite directions. CogVideoX-3, the current generation of the line the company has developed since 2024, ships no weights at all: it is sold through the API and nothing is available to download. SCAIL-2, the character animation model the company released as open weights, does not run on that line. It runs on Alibaba's. The evidence is not subtle. The download bundles Alibaba's Wan 2.1 image encoder and Google's umT5-XXL text encoder, the inference command loads a checkpoint named SCAIL-14B, and the default branch of the code repository is called wan-scail2. There is a second branch, sat-scail2, for the company's own training stack, but the weights that were actually published are the ones for the competitor's runtime. The training material came the same way. Z.ai states that 60,000 motion pairs were synthesised using three off-the-shelf models, two of them from other labs: Alibaba's Wan-Animate and Orange's MoCha. The company calls the method reverse driving and reports that the finished model acquired behaviour its teachers never had — animal-driven motion, cross-identity replacement, and control from mesh renders it was never trained on. One detail is worth checking before anyone redistributes the file. The code repository is Apache 2.0. The weights repository declares MIT in its metadata and contains no licence file at all, while the archive it hands over includes third-party components released under Apache 2.0, a licence that asks for its notices to travel with the code. The badge and the contents do not say the same thing. SCAIL-2 now has a profile in our catalogue, with the licensing situation recorded as the publisher left it.
SCAIL-2 →Z.ai sells its speech model for meeting minutes and live captions. The API accepts 30 seconds of audio.
Two pages of Z.ai's own documentation, checked on 3 September 2026, describe the same speech-recognition model in ways that cannot both be acted on. The model card for GLM-ASR-2512 lists six uses. The first is real-time meeting minutes — "transcribe online meetings instantly, automatically organizing structured summaries". The third is live video captioning for news broadcasts, courses and video conferences. The sixth is medical record entry: a doctor dictates a patient history, the model writes it down. The API reference for the same model sets the ceiling: supported formats .wav and .mp3, file size 25 MB or less, and audio duration 30 seconds or less. Per call. The limit is not a side effect of the non-streaming path. Z.ai documents a streaming mode for this endpoint, and the same specification applies to it. Streaming here means the transcript comes back progressively while the model works — not that audio can be fed in continuously. There is no long-form endpoint, no asynchronous job for a recording, and no documented way to hand over an hour of audio. What that leaves the caller is arithmetic. A one-hour meeting is 120 separate requests, each cut at a fixed half-minute boundary that lands wherever it lands: mid-word, mid-sentence, mid-name. Every seam is a place where context is lost and where the custom dictionary — one of the model's genuine strengths — has to be resupplied. Speaker continuity across the joins is the caller's problem, not the model's. None of this makes GLM-ASR-2512 a weak model. Z.ai quotes a character error rate of 0.0717, the language coverage is unusually broad for the price — Sichuanese, Cantonese, Min Nan and Wu alongside Mandarin, English and dozens of other languages — and at roughly $0.0024 per minute it undercuts OpenAI's currently recommended transcription model by about half. For short utterances, voice input, commands and single customer-service turns, the shape fits. It is the advertised use cases that do not fit, and the gap is worth naming because of who it catches. A developer reading the model card sees meeting transcription promised in the first line and only meets the 30-second wall after writing the integration. There is an ending that is unusual for a paid product. Z.ai published an open sibling on the same day, GLM-ASR-Nano-2512: 1.5 billion parameters, MIT licence, free to download, and — because it runs on the caller's own hardware — subject to no per-call duration cap at all. For the three long-form uses the vendor advertises, the free model is the one that can actually do them.
GLM-ASR-2512 →Reasoning stopped being a product. It became a setting.
Two years ago, buying a model that thinks before it answers meant buying a different model. Today, at four of the five largest publishers, it means passing a parameter — and the separate reasoning lines are being switched off one by one. OpenAI's deprecation page now carries an end date for every surviving member of the o-series. o1, o1-pro, o3-mini and o4-mini shut down on 23 October 2026; o3 and o3-pro follow on 11 December. The recommended replacement in each case is a general model from the GPT-5.6 family, and for the two pro-tier models the replacement is written out as gpt-5.6-sol with reasoning mode set to pro. The premium reasoning tier is now literally a value in a request. Mistral got there first. Its entire Magistral line — the company's only reasoning products — was retired by 31 July 2026, and the migration targets listed by the manufacturer were not further reasoning models but ordinary models of the main line. DeepSeek reached the same place by a different route. Its current price list contains three models, all of them V4, and a single row headed thinking mode: both non-thinking and thinking modes are supported, with thinking as the default. The R1 line that made the company's name in January 2025 has no entry of its own. Alibaba's separate line ended earliest of all, and quietly. QwQ-32B-Preview appeared in November 2024 and QwQ-32B in March 2025; no third model ever followed. From Qwen3 onwards, thinking is a switch inside the general model. QwQ has a profile in this catalogue as of today — as the closing entry of a product category rather than the first entry of a family. Google is the exception that clarifies the rule, because it never made the separation in the first place. Gemini shipped thinking as a budget the caller sets, not as a model you choose, and so has nothing to wind down. For readers the practical consequence is narrow but real. Where a hosted reasoning model is switched off, its published benchmark results become unreproducible — the closed Magistral Medium 1.2 is already in that position. Where the weights were open, the model survives as a download after it stops being a product: QwQ-32B is still pulled tens of thousands of times a month, eighteen months after the line it belongs to quietly ended.
QwQ-32B →Alibaba's new animation model was built for live video — and the live-video version is the one part it did not release
Alibaba published Wan2.2-Animate-2-14B on 7 August 2026 under Apache 2.0: weights, inference code and, a day later, a paper. It animates a still character from a driving video of a person performing, and the engineering behind it is a genuine redesign rather than a refresh. The first generation extracted a pose skeleton from the driving video and handed that to the generator; the team's own paper names what goes wrong with that arrangement, in plain terms — extraction errors travel into the output, and the character's identity drifts as the clip runs. Animate-2 removes the extractor and feeds the driving video straight into a rebuilt diffusion transformer. The paper's own framing, though, is not about fidelity. It opens by pointing out that every existing system of this kind is designed for offline synthesis and cannot meet the requirements of interactive applications such as digital avatars and live-streaming hosts. Its answer is Wan-Animate-2-Lite, an efficient variant trained in three stages — teacher-forcing pretraining with an error buffer, then Self-Forcing distillation with chunk-wise backpropagation — to pull latency down to real-time thresholds. That is the claim the abstract ends on. Lite is not in the repository. What is there are two checkpoints of identical size, 32.8 GB each: a base build that runs 40 denoising steps with classifier-free guidance, and a distilled build that reaches a picture in 10 steps with guidance off. Distillation makes the offline model roughly four times cheaper to sample; it is not the streaming architecture the paper describes. And the omission does not read as a queue: the to-do list on the model card has every item ticked — inference code, checkpoints, Diffusers, DiffSynth-Studio, ComfyUI — with no line for Lite at all. A smaller detail points the same way. The citation block at the bottom of the card still carries an unfilled placeholder where the arXiv number should be, while the correct number sits in a badge at the top of the same file. This was shipped quickly. What readers can act on is the ordinary part. The download counters show demand pointing at speed rather than quality: on the Diffusers conversions published the day before the release, the distilled build is pulled around three times more often than the full-quality one. The first-generation Animate, meanwhile, is still being fetched an order of magnitude more than either — which is what usually happens in the first month of a model that changes how it is driven. The base repository reports no downloads at all against 185 bookmarks, but that counter tends to read zero for repositories shipping raw checkpoint files rather than a standard configuration, so it says more about how the files are pulled than about interest in them.
Wan2.2-Animate-2-14B →