News
What's happening in robotics and AI — curated by the wujec.ai editors.
DeepSeek's new model beats Opus and GPT on the agent test everyone has already saturated. On the two harder versions of that same test, it loses by twenty points
DeepSeek published V4.1-Flash on 10 September 2026 with a comparison table that reads like a clean sweep — until you follow one test through three of its versions. On Terminal-Bench 2.1 the new model scores 90.6, ahead of Opus-5.0 at 89.1 and GPT-5.6 Sol at 88.8. On Terminal-Bench 3.0 it scores 30.0 against Opus-5.0's 43.3. On version 4.0, 31.2 against 51.8. The three numbers describe the same model on the same family of tasks, and the spread is a lesson in how to read a benchmark. Version 2.1 has been in circulation long enough for every serious laboratory to score in the high eighties; a 1.5-point lead there separates models that are, in practice, equally capable. Versions 3.0 and 4.0 are the harder revisions written precisely because the old one stopped telling models apart — and there the distance between DeepSeek's model and the most expensive Western one is roughly twenty points. The same pattern repeats elsewhere in the table. On Humanity's Last Exam, a test still far from saturated, V4.1-Flash reaches 36.8 against Opus-5.0's 56.3. On ProgramBench, 20.3 against 37.0. What the model does win is worth stating plainly, because it is not a small thing. Its Codeforces rating of 3,471 is the highest in the table. It resolves 74.2 percent of DeepSWE v1.1 tasks, leads on CyberGym with 88.1 and on AutomationBench with 54.8. And it does this while activating 8 billion parameters per token when reading a prompt and 16 billion when writing — out of 552 billion in the backbone. The efficiency is the actual headline. DeepSeek rebuilt the part of the model that stores the conversation, cutting the memory held per token to 890 bytes, about a quarter of what the July model needed. Through the company's API the new model costs $0.30 per million input tokens at peak and $1.20 per million output, against $0.44 and $1.32 for its predecessor, with off-peak hours at half price. Weights are MIT-licensed and were downloaded 140,636 times in the following weeks. Read together, the table says something more useful than 'best model'. For agent work at scale, where a million-token session runs on a budget, this is a remarkable instrument at a fraction of frontier pricing. For the hardest reasoning a buyer can currently pose, the expensive models remain ahead, and DeepSeek's own numbers say so.
DeepSeek-V4.1-Flash →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 →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 →Every Wan 2.2 model card points readers at a licence file the repository does not contain
Alibaba's Wan 2.2 video models all close their model cards with the same paragraph: the weights are Apache 2.0, the company claims no rights over what users generate, and "for a complete list of restrictions and details regarding your rights, please refer to the full text of the license" — with the phrase "license" linking to a file named LICENSE.txt inside the repository. That file is not there. We checked all four Wan 2.2 weight repositories on 2 September 2026: T2V-A14B, I2V-A14B, S2V-14B and Animate-14B each return 404 for LICENSE.txt. The two Wan 2.1 repositories we checked, T2V-14B and T2V-1.3B, do carry the file, with the standard Apache 2.0 text in full. The practical effect is smaller than it sounds, but it is not nothing. Two independent declarations of Apache 2.0 remain in place: the licence tag in the repository metadata on Hugging Face, which is what the platform displays and what automated licence scanners read, and the LICENSE file in Alibaba's matching code repository on GitHub. Anyone shipping a product on these weights has a licence. What they do not have is the document the model card tells them to go and read — the one it says contains the complete list of restrictions. The timing gives this more weight than a filing error usually deserves. Wan 2.2 is the last generation of Alibaba's video line whose weights were published at all; everything since — 2.5, 2.6, 2.7 and Wan3.0 — is served through an API only. These four repositories are therefore not an archive. They are the files anyone who wants a Wan they can run on their own hardware downloads today, and one of them, Animate-14B, is the most bookmarked video model the company has ever released.
Wan2.2-Animate-14B →Mistral no longer sells a reasoning model: every Magistral now sits in the retirement table
In June 2025 Mistral AI launched Magistral, its own family of models that write out a chain of reasoning before answering. Fifteen months later the family is gone from the vendor's API. All six members — Magistral Small 1.0, 1.1 and 1.2, and Magistral Medium 1.0, 1.1 and 1.2 — appear in the retirement table of the lab's own documentation. The last two were switched off on 31 July 2026. The replacements named in that table are not another reasoning model. Users of Magistral Small are pointed to Mistral Small 4, users of Magistral Medium to Mistral Medium 3.5 — both ordinary mainline models. Reasoning did not disappear from the lab's line-up; it stopped being a product of its own. Since December 2025 the Ministral 3 family has shipped in three flavours of every size — base, instruct and reasoning — so a customer who wants deliberation picks a variant of the standard model instead of a different model. That is a pattern worth naming, because Mistral is not alone in it. A separate reasoning line made sense while the technique was new and expensive enough to be sold at a premium. Once it became a training stage that any model can be put through, keeping a parallel family meant maintaining two catalogues of the same sizes. What makes the case unusual is that switching a model off in the API did not switch it off in the world. Magistral Small 1.0, retired from the vendor's service on 30 November 2025, was downloaded roughly 80,700 times from Hugging Face in the thirty days to early September 2026 — more than six times the traffic of the final version 1.2, which is technically the better model. Devstral Small 2 makes the same point harder: withdrawn from the API on 31 March 2026, it pulled about 229,700 downloads in that same window, twelve times more than the 123-billion-parameter flagship it was released alongside. The counters cover thirty days only, and older releases have an advantage — tutorials, quantised forks and pipelines that were written once and still point at them. But the direction is clear enough. For a lab that publishes weights under Apache 2.0, retirement is a statement about what the company is willing to host, not about what people are running. That is why our profiles of Magistral Small 1.2 and Devstral Small 2, added today, are marked historical and at the same time describe how to run them — the service ended, the software did not.
Magistral Small 1.2 →Z.ai opened the weights of its cheap model — the flagship it promised two weeks ago is still closed
Z.ai published GLM-5.3-Flash on 26 August 2026 and put the weights on Hugging Face the same week, under a plain MIT licence. Two weeks earlier the company launched its flagship GLM-5.3 with a promise that its weights would follow in about a fortnight. That deadline has now passed, and the flagship repository does not exist: the newest Z.ai model anyone can download is the cheap one. The last flagship whose weights were actually released is GLM-5.2 from June 2026, at 753.3 billion parameters. GLM-5.3-Flash is less than half that size — 320 billion parameters with 18 billion active per token, and the published safetensors index confirms it at 321.3 billion. The smaller model is not merely a trimmed version. Z.ai describes it as the first open-weight frontier model to combine sparse attention with linear attention, and quantifies the gain against GLM-5.3: attention computation down 3.01 times, KV cache down 4.44 times. It is also the first natively multimodal model in the family, meaning it can look at a rendered interface and correct its own code rather than working blind. The price is where that architecture shows. GLM-5.3-Flash lists at 0.15 dollars per million input tokens and 0.50 per million output; GLM-5.3 costs 1.40 and 4.40. That is 9.3 times cheaper on input and 8.8 times cheaper on output — close to, but not quite, the one tenth of the price the company claims. A 50 percent promotion runs until 9 September 2026, which for now doubles the gap again. The performance claims stay in the maker’s own hands. Z.ai says the model beats GLM-5.2 across benchmarks and approaches Claude Opus 4.8 on coding and agentic tests, but publishes those results only as a chart image, and no independent laboratory has repeated them. What can be verified is the licence: the LICENSE file in the repository is the bare MIT text, with no attribution rider and no restriction on commercial use.
GLM-5.3-Flash →Open weights stopped asking permission — of the forty most downloaded models, only Meta's still need an account
We read the licence of every repository in the forty most downloaded text-generation models on Hugging Face today, 25 August 2026, and checked whether the weights can actually be fetched. Two of the forty are gated: Meta's Llama-3.2-1B-Instruct and Llama-3.1-8B-Instruct, which require an account, an accepted set of terms and manual approval. A third, Google's Gemma 3 1B, is gated for the same reason. Everything else in the list downloads without an account. The change that made this true is recent and belongs to Google. Gemma 3, published under Google's own Gemma Terms of Use, could not be downloaded — or even have its licence text read — without logging in; a request for the file returns HTTP 401. Gemma 4, published on 11 March 2026, is under plain Apache 2.0 and is not gated at all. It is not a marginal model: the 31B instruction-tuned repository alone records 8.79 million downloads in thirty days, more than any other frontier-class open-weight model we track. What remains of bespoke licensing is narrower than its reputation. Moonshot's Kimi K2 carries a "Modified MIT" licence whose only modification is a display duty: a product with more than 100 million monthly active users or more than 20 million dollars in monthly revenue must show the words "Kimi K2" in its interface. DeepSeek-V3, the model on which that lab's whole line stands, is still routinely described as MIT-licensed and is not — its repository ships MIT for the code and a separate DeepSeek License Agreement for the weights. Only from March 2025 did the company put the weights themselves under MIT. One habit is worth flagging for anyone who takes a label at face value. Several heavily downloaded repositories declare a licence in the card metadata and ship no licence file at all, and quantised repacks inherit the label from a model card rather than from the original terms. The label is metadata; the file is the contract. Where the two disagree, only one of them is enforceable.
Gemma 4 31B →MiniMax's most downloaded model is the one you may not sell anything with
MiniMax M2.7 is pulled from Hugging Face roughly 909,000 times a month — more than any other text model the Chinese lab has published, and more than four times the traffic of its own successor M3. It is also the only one in the family that forbids commercial use. The licence file shipped with the weights is titled "Non-Commercial License". Personal use, self-hosted deployment, research and experimentation are expressly free of charge. Everything else is not: selling a product or service that relies on the model, putting it behind a paid API, or deploying a fine-tuned derivative for commercial purposes all require prior written authorisation from MiniMax, requested by e-mail. Anyone who obtains that permission must also display "Built with MiniMax M2.7" on the product. That term is an outlier rather than a direction of travel. MiniMax M2 (October 2025) and M2.1 (December 2025) were released under MIT with a single added clause: a commercial product built on the model must show the model name in its interface. In M2 that duty starts only above 100 million monthly active users or $30 million in annual recurring revenue; in M2.1 it applies with no threshold at all. M2.5 (February 2026) moved to the company's own MiniMax Model License, which still allows redistribution but requires a fixed attribution notice to accompany every copy. M2.7, released a month later, closed commercial use altogether. M3, in June 2026, reopened it under a community licence that permits commercial deployment in exchange for visible attribution. The practical consequence is easy to miss, because nothing about the download page signals it. Four models in this line share the same body — 62 layers, 256 experts, eight routed per token, and 228,689,764,864 parameters in M2, M2.1 and M2.7, to the byte. A team that swaps one for another as a drop-in upgrade changes its legal position without changing a line of code. The strongest freely reusable model in the line is M2.5; the strongest one overall, by the producer's own benchmark table, is the one that needs a signature. wujec.ai has today added catalogue profiles for M2.1, M2.5 and M2.7, each stating the licence terms in full.
MiniMax M2.7 →Mistral built its name on freely licensed weights — its biggest one is still the model it published in April 2024
Mistral AI became a recognised name in September 2023 by publishing a 7.3-billion-parameter model under Apache 2.0 — the licence that lets anyone download, fine-tune and commercially resell the result without asking. Three months later, Mixtral 8x7B repeated the trick at a size that mattered: 46.7 billion parameters stored, 12.9 billion spent per token, benchmark scores at the level of Llama 2 70B and of the GPT-3.5 base model then serving free ChatGPT. In April 2024, Mixtral 8x22B scaled the same sparse design to 141 billion parameters, 39 billion of them active per token, with a 64,000-token window and native function calling. It, too, went out under Apache 2.0. That April 2024 release is still the largest model Mistral has ever opened. Everything above it changed terms: Mistral Large 2.1, from November 2024, shipped its weights under the Mistral Research License with a separate commercial licence sold on top, and the flagships that followed were never opened the same way. The company that made permissive licensing its signature has kept publishing open weights — but at the small and medium end, while the top of its range moved behind commercial terms. The practical consequence is that Apache 2.0 cannot be revoked. Mixtral 8x22B has disappeared from Mistral's own API line-up, which today lists Large 3, Medium 3.5, Small 4 and Ministral 3, and the company has never announced a shutdown date for it. That is not a problem for anyone relying on it: the weights sit on Hugging Face, third-party providers still serve them, and a licence granted under Apache 2.0 stays granted regardless of what the vendor's catalogue says. wujec.ai has now catalogued all three of these open releases.
Mixtral 8x22B →Alibaba's downloadable flagship is not the flagship it rents — the model card lists four things it cannot do
Qwen3.8-Max, the 2.446-trillion-parameter model Alibaba published for download on 8 August, is the first Max-class model the company has ever released as open weights. What almost no coverage mentions is that the file you download and the model you call through the API are not the same product — and the difference is written by Alibaba itself, in a note near the top of the model card. The note says the API version is "based on" the released checkpoint "with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools". Read as a list of what the download lacks, it comes to four items: - **No vision.** The repository ships a text-only architecture (`qwen3_5_moe_text`, pipeline tag `text-generation`). The hosted Qwen3.8-Max accepts images and video; the download does not. - **No way to switch reasoning off.** The open checkpoint always reasons before answering — depth is adjustable through `reasoning_effort`, but the non-thinking mode exists only in the API. - **262,144 tokens, not a million.** The card states the context length plainly: 262,144 natively, extensible to 1,010,000. The hosted version has a million by default; self-hosters have to extend it themselves and accept whatever that does to quality. - **No built-in tools.** Web search, code execution and the rest of the hosted toolchain are part of the service, not the weights. None of this is concealed — it is one paragraph in Alibaba's own card, which is precisely why it is worth repeating: the benchmark numbers circulating for Qwen3.8-Max (Terminal-Bench 2.1 at 86.6, SWE-bench Pro at 67.7, DeepSWE 1.1 at 56.6) were measured on the hosted configuration, and a self-hosted deployment starts from a narrower one. The card also settles what the model actually is architecturally, which the launch materials left vague. It is a hybrid: 92 layers arranged as 23 repetitions of three Gated DeltaNet blocks followed by one gated attention block, so only a quarter of the layers use full attention and the rest use linear attention. The mixture of experts holds 512 experts, of which 10 routed plus 1 shared fire per token. That is the design that makes 2.4 trillion parameters cost 95 billion per token to run — and it is the first time Qwen has shipped it at flagship scale. Our profile of the model now carries the native context figure, the hybrid architecture and an explicit note that the published weights are text-only.
Qwen3.8-Max →Meta promised its flagship weights a week ago — its download page still holds only the distilled model
On 10 August 2026 Mark Zuckerberg published a long essay arguing that American labs should lead the open-weight movement, and said Meta would open the weights of Muse Spark 1.2, its most capable model. The word he used for the timing was "soon". No date was given. Seven days later the company's model account on Hugging Face holds four repositories, and all four are the same smaller model. Muse-Glimmer-30B, the base weights, went up on 9 August. Next to it sit a GGUF conversion for local runtimes, a draft head for speculative decoding published as an "assistant" repository, and an ExecuTorch build for mobile deployment. There is no Muse Spark repository of any kind — not the 1.2 release, not the 1.1 one that preceded it in July. The distinction matters more than it may look, because the two models are not alternatives. Meta describes Muse Glimmer as a distillation of Muse Spark: a roughly 29.6-billion-parameter student trained on the teacher's outputs, with a 131,072-token context window against the flagship's million. What can be downloaded today is the compressed derivative of the model whose weights were promised, not the model itself. Demand for the derivative has been considerable. Meta's own four repositories record 780,323 downloads in the eight days since publication — 334,099 for the base weights, 395,175 for the GGUF conversion, 43,909 for the drafter and 7,140 for the mobile build. A single community quantisation published by Unsloth adds another 755,125, taking the family past 1.5 million downloads in little over a week. All of it carries a plain Apache 2.0 licence with no revenue or user thresholds, which is itself a departure from the bespoke community licences Meta attached to its earlier open releases. Until the flagship weights appear, the practical position is unchanged: Muse Spark 1.2 remains reachable only through Meta's interfaces, and the open-weight commitment stands as an announcement rather than a file. wujec.ai will update the Muse Spark 1.2 profile on the day a repository appears.
Muse Spark 1.2 →Qwen's three billion downloads led the headlines — the report behind them says the giants are 1% of the traffic
Hugging Face published its half-yearly stocktake of open-weight models on 14 August 2026, and by the next morning one number from it was everywhere: Alibaba's Qwen family has passed three billion downloads, ahead of Google and Meta combined. The figure is Alibaba's own cumulative count across all platforms. The report's own measurement is narrower and more useful: on Hugging Face alone, Qwen repositories were downloaded 2,045 million times during 2026, against 418 million for Google and 227 million for Meta. Qwen now carries more than half of all open-model downloads on the hub, and its models have been forked into 151,448 derivatives — roughly 180 to 210 new ones every day, against 82,506 for Google and 32,155 for Llama. The finding that went unreported sits a few charts further down, and it undercuts the way this race is usually described. Models below one billion parameters account for 83% of all downloads ever recorded on the hub. Models above 100 billion parameters — the frontier, the ones that get the launch events and the benchmark tables — account for 1%. Restrict the count to 2026 alone and everything above 70 billion parameters is still only 3% of the volume. Distribution is brutally concentrated in another direction too: 85.6% of all models on the hub have fewer than 200 lifetime downloads, while 1.5% of them take 99.2% of the traffic. The same asymmetry shows up in what people actually run at home. In GGUF, the quantised format used for local inference on ordinary hardware, Qwen is pulled 39.6 million times a month, Google's Gemma 20.8 million and Llama 7.5 million — the last despite Llama having more GGUF repositories published than Qwen. The report reads this as a strategy difference rather than a quality verdict: Qwen ships across every size class, from sub-billion models that run on a laptop up to the frontier, while a lab that publishes only large models collects a fraction of the traffic. Moonshot, whose open portfolio is frontier-only, recorded 37 million downloads — 55 times less than Qwen. Two further findings are worth recording for anyone tracking who is publishing what. First, licensing: among Chinese releases above 20 billion parameters, 59% carry Apache 2.0 and 22% MIT, and none carry a non-commercial restriction — which is the backdrop to the licensing split we described in Qwen3.8 yesterday, where the small model got Apache 2.0 and the large one did not. Second, authorship at the top end: the report notes that most US releases above 100 billion parameters are derivatives of Chinese base models, and that the largest volume of new repositories now comes from hardware vendors — AMD and NVIDIA, with more than 200 new repositories each — publishing optimisation and conversion layers rather than original models. The hub itself grew from 2.43 million to 2.96 million models over the period.
Qwen 3 →Alibaba opens both ends of Qwen3.8 — but only the small model gets Apache 2.0
Alibaba has now published downloadable weights for both halves of the Qwen3.8 generation, and the licences attached to them are not the same document. Qwen3.8-27B, released on 14 August, is a dense 27.78-billion-parameter vision-language model under a plain Apache 2.0 licence — no thresholds, no conditions, no separate agreement for commercial use. It takes text, images and video, has a native 262,144-token context that Alibaba says stretches to a million, and runs on Transformers, vLLM, SGLang and TokenSpeed. Alibaba's own benchmark table puts it at 61.7 on SWE-bench Pro against 57.6 for its closed Qwen3.7-Plus, and at 42.2 on DeepSWE 1.1 against 13.3 for the previous 27B generation — vendor-reported figures, not independently reproduced. The flagship went out under different terms. Qwen3.8-Max, 2.446 trillion parameters, is on Hugging Face as Qwen3.8-2.4T-A95B in bf16 and FP8, but under a bespoke "Qwen3.8-Max License". Its text reads like a permissive licence for almost everyone: free use, modification, hosting, fine-tuning and resale. Two clauses change that at scale. A product with more than 100 million monthly active users or 20 million dollars in monthly revenue must display the model's name prominently in its interface. And any company running a model-as-a-service or an AI coding/office assistant business whose revenue passes 50 million dollars over twelve months must obtain a separate licence from Qwen before commercial use — internal deployment excepted. The practical effect is a licence aimed squarely at a handful of named competitors while leaving everyone below those thresholds fully free. It is a pattern worth watching: "open weights" is quietly becoming a spectrum, and the difference between the two Qwen3.8 downloads is not the code but the paragraph that comes with it. Both models are now in the catalogue, with the licence terms written out in full in the specification tables.
Qwen3.8-27B →DeepSeek ships the final V4-Pro: open weights, 1M context, agentic scores up sharply
DeepSeek published DeepSeek-V4-Pro-0813 on Hugging Face on 13 August 2026, describing it as the official release of V4-Pro and the successor to the preview weights that had carried the name since April. The repository went up under the MIT licence, so the full model — 1.65 trillion parameters in a sparse mixture-of-experts layout, 61 layers, 384 routed experts with six active per token, shipped in FP8 — can be downloaded and self-hosted. The context window stays at 1,048,576 tokens. The architecture is unchanged from the preview; what is new is a DSpark speculative-decoding module attached to the model and a reworked reasoning control. The `reasoning_effort` parameter now takes three levels — low, high and max — letting callers trade deliberation time against cost. DeepSeek also dropped the Jinja chat template in favour of a documented Python encoder, a change that will require work from anyone running the weights locally. The gains DeepSeek reports are concentrated in agentic and coding work, and they are large. On the company's own table the new build reaches 62.7% on DeepSWE against 12.8% for the preview, 61.5% on NL2Repo against 38.5%, 83.3% on Cybergym against 52.7% and 87.9 on Terminal Bench 2.1 against 72.1. On Humanity's Last Exam it scores 42.7% without tools and 60.0% with them, up from 37.7%. Against rivals in the same table the picture is closer. Kimi K3 edges it on Terminal Bench (88.3) and DeepSWE (67.5), and Anthropic's Fable 5 leads on Humanity's Last Exam (53.3 without tools). What DeepSeek keeps is the position it has held all year: comparable results at frontier level, with the weights published rather than rented. Benchmark figures here are the vendor's own and have not yet been reproduced independently.
DeepSeek-V4-Pro →The model Europe was not allowed to have: Tencent drops the regional carve-outs from Hy3 and goes global
Chinese open-weight models have spent the past two years competing on size and price. Tencent's Hy3 has quietly been competing on something else: who is allowed to download it. When the preview build appeared in April 2026, its licence carried a geographic exclusion — the European Union, the United Kingdom and South Korea were carved out. The official release on 6 July 2026 removed the clause entirely and put the weights under a plain Apache 2.0 licence, one of the most permissive terms yet attached to a model of this scale. On 5 August Tencent completed the move by announcing global availability: the WorkBuddy assistant, the Miora design tool and Tencent Cloud TokenHub, plus third-party access through OpenRouter and downloads on Hugging Face and ModelScope. The model itself is a sparse mixture-of-experts design: 295 billion parameters in total, roughly 21 billion activated per token across 192 experts with top-8 routing, 80 layers, and a separate 3.8-billion-parameter multi-token-prediction stack that speeds up decoding. The context window is 256,000 tokens. The distinctive feature is that the amount of deliberation is a setting rather than a separate product — no_think answers immediately, think_low reasons briefly, think_high runs an extended chain — so one endpoint covers both a chat reply and a long analytical job. Tencent's own framing is that Hy3 performs comparably to flagship models with two to five times as many parameters, across reasoning, instruction following, in-context learning, code generation and agentic work. The company also reports that within a week of launch, API traffic to Hy3 ran 68 times higher than to the previous generation, putting it first on OpenRouter's usage leaderboard. Listed OpenRouter pricing is 0.1288 US dollars per million input tokens and 0.5336 per million output; access through WorkBuddy is free until 31 August 2026. The usage number is the one to watch. Benchmark parity claims from vendors are routine; a 68-fold jump in paid API calls is a statement about how many developers actually switched. Independent coverage broadly agrees with the performance claims, with coding the one area where rival open-weight models still hold an edge.
Tencent Hy3 →DeepMind opens its cyclone forecaster: three-day tracks as good as yesterday's two-day ones
Google DeepMind has published WeatherNext Cyclones in Nature and released the model weights on GitHub, alongside the broader WeatherNext 2 system. The headline claim is a full day of lead time: three-day forecasts of a storm's track and intensity now reach the accuracy that previous models achieved only at two days. The model was trained on roughly 20 terabytes of atmospheric data and about 5,000 historical storms, and it forecasts by ensemble — DeepMind scaled the run from 50 scenarios to 1,000, which is what allows a forecaster to read the spread as a probability rather than a single line on a map. A 15-day forecast at 28 by 28 kilometre resolution runs in under a minute on a single TPU. For a catalogue of AI models this is a useful reference point on what open weights now cover. Weather prediction has been a showcase for machine learning for several years, but the operational systems behind national forecasts have stayed closed. Releasing the weights moves a model of this class into the hands of meteorological services that cannot afford to train one. Two details fill out the picture. The first is who checked the work: DeepMind says the model was developed with the US National Hurricane Center, the Cooperative Institute for Research in the Atmosphere and the UK Met Office, and points to the 2025 season — Hurricane Melissa’s rapid intensification and its landfall on Jamaica — as a case of operational use. The second is intensity, historically the harder half of the problem and the half that decides evacuation orders: the model reports around 11 knots of error, better than the specialised HWRF hurricane model, while track error at day three sits near 100 km, ahead of the ECMWF ensemble. That it does this on a 28 by 28 kilometre grid is the awkward part for the traditional approach. The cells are about a hundred times coarser than the physics-based models it is measured against, and a hurricane eye fits inside one of them. The model is not resolving the storm at all; it has learned what storms of a given shape and history tend to do next.
WeatherNext Cyclones →NVIDIA's open robot brain gets a production licence: GR00T N1.7 reaches general availability
NVIDIA's GR00T line of open robot foundation models has quietly passed the point that matters most to anyone trying to build a product on it. When N1.7 was announced in early access on 17 April 2026, the headline item was not a benchmark but a legal clause: the release ships, in NVIDIA's own words, with "a license that supports production use cases". Earlier versions of the model were distributed under a non-commercial NVIDIA licence. The Isaac-GR00T repository now describes N1.7 as generally available, with support and stability guarantees; NVIDIA has not attached a date to that step, so we report the general availability as a state of the repository rather than as a dated event. Under the licence change sits a real architectural swap. The vision-language backbone is now Cosmos-Reason2-2B, built on the Qwen3-VL architecture, replacing the Eagle backbone used in N1.6 — NVIDIA's own reasoning model has moved inside NVIDIA's own robot policy. The action head remains a flow-matching diffusion transformer, but its depth reverses the previous release: N1.6 had doubled it from 16 to 32 layers, and N1.7 cuts it back to 16. Two capabilities are new: structured reasoning at task and subtask level, and finger-level actuation aimed at contact-rich work such as small-parts assembly. The model card is unusually forthcoming about cost. Pretraining used 21.6 million data points from 13 datasets, and NVIDIA publishes what that run consumed: 41,288 kWh and 16.949 tonnes of CO2 equivalent. The deployment numbers are equally concrete, and they show the price of putting the policy on the robot itself. With TensorRT on an H100 a step takes 27.9 ms — 35.9 Hz. On a Jetson Orin only the diffusion transformer can be compiled, because TensorRT 10.3 does not support the backbone engine, and the same step takes 216.5 ms, or 4.6 Hz. One inconsistency is worth flagging, because both statements come from NVIDIA: the research page for N1.6 described that version as using an internal Cosmos-2B variant, while the N1.7 repository states that N1.6 used the Eagle backbone. We record both and prefer neither. Older versions have not been withdrawn — N1.6 and N1.5 remain available in the n1d6 and n1d5 branches. GR00T N1.7 now has its own profile in this catalogue, alongside the original N1 from March 2025. This item runs without an illustration. NVIDIA's logo is not a plain wordmark, and we could not establish a licence that permits reuse without attribution, which our news illustrations currently require.
NVIDIA Isaac GR00T N1.7 →NVIDIA hands the driving model away: Alpamayo 2 Super comes with a licence to sell what you build on it
NVIDIA has released Alpamayo 2 Super, a vision-language-action foundation model for autonomous vehicles, under the Linux Foundation's OpenMDW-1.1 licence — which permits fine-tuning, derivative models and commercial redistribution. For a category where driving stacks are usually the most closely guarded asset a company owns, that licence is the news as much as the model is. The model pairs a 32-billion-parameter Cosmos 3 Super Reasoner with a 2-billion-parameter diffusion action expert, 34 billion parameters in total according to the published model card. It does not only emit a trajectory: it also returns a chain-of-causation trace explaining why it chose that trajectory, a meta-action, automatic labels and grounded visual question answering. On the LingoQA driving-comprehension benchmark NVIDIA reports it first among roughly forty models, 17.0 points ahead of Qwen2.5-VL 72B and 15.1 ahead of Gemini 2.5 Pro. One caveat for readers comparing numbers: NVIDIA's own blog post rounds the model to "30 billion" parameters, while the model card and repository give 34 billion (32B plus 2B). We use the model-card figure.
Mistral opens a 3B moderator that reads the rulebook at run time — not during training
Mistral AI released Shieldstral 1.0 on 4 August 2026: a 3-billion-parameter safety classifier published under Apache 2.0, the first moderation model the company has given away with open weights after two hosted APIs. The design choice is the story. Guard models usually ship with a fixed taxonomy of harms decided during training, so a platform whose rules differ from that taxonomy — or whose rules change — is stuck between retraining and mismatch. Shieldstral turns moderation into a single yes/no question asked at inference time: the developer supplies the context, a plain-language policy question and the content, and the model returns only the probabilities of "yes" and "no", normalised into a continuous score. Changing a policy means editing a prompt. The reported numbers put a small model against much larger ones. Mistral claims 88.1 F1 on WildGuardTest, 84.1 on ToxicChat and 99.4 on HarmBench prompt classification, plus 97.7 on the multimodal VLGuard set — matching or beating open guard models up to seven times its size, on the vendor's own evaluations. It covers twelve languages, judges images as well as text, and was trained at a 32k context. The practical figure is 16 GB. That is a single mid-range GPU, and with llama.cpp quantisation it drops to a CPU. A moderation layer that fits on one machine is available to operators who legally or commercially cannot ship user content to a third-party API — hospitals, public administration, anyone under strict data-residency rules. It is also, conveniently for Mistral, the argument European buyers have been asking for. One limitation worth stating plainly: every benchmark figure above comes from the company that built the model, and guard models are notoriously sensitive to how the policy is phrased. The weights are public, so independent numbers should follow.
Shieldstral 1.0 3B →Alibaba launches Qwen3.8-Max, a 2.4-trillion-parameter flagship
Alibaba released Qwen3.8-Max on 3 August 2026, the largest model the Qwen family has produced so far. It is a sparse mixture-of-experts design: 2.4 trillion parameters in total, of which roughly 95 billion are activated per token, with a context window of up to one million tokens. The model is available worldwide through Alibaba Cloud's Model Studio APIs and through QwenWork, the company's workplace agent platform. Alibaba said full model weights would follow for public download, alongside a smaller Qwen3.8-27B variant aimed at hardware-constrained deployments. On the Arena.AI leaderboard Qwen3.8-Max became the highest-ranked Chinese model for text tasks and placed second globally for vision, putting it in the same bracket as current frontier systems from OpenAI and Anthropic. Alibaba shares rose sharply in Hong Kong on the announcement.
Qwen3.8-Max →Cisco's 1B-parameter bug hunter: Antares localizes vulnerabilities without your code ever leaving the building
Cisco's Foundation AI group has released Antares-350M and Antares-1B, two open-weight small language models built for a single security task, and published them on Hugging Face. The task is vulnerability localization: given a codebase and the description of a known flaw, find the files that actually contain it. It sounds narrow, and that is the point. Security teams triaging advisories spend much of their time answering exactly this question across repositories they did not write, and frontier models are an expensive instrument for it. What Antares does differently is work the problem the way a human investigator would rather than scanning against fixed rules — starting from the vulnerability description, searching for relevant code patterns, reading candidate files, folding new evidence back into the search and abandoning paths that lead nowhere. Cisco's claim is that a model shaped around that loop does not need to be large. The comparisons the company publishes are pointed. Antares-1B is said to outperform Google's Gemini 3 Pro on this task and to sit level with Z.ai's GLM-5.2 — a 744-billion-parameter model. Cisco also previews an unreleased Antares-3B that it says beats both GLM-5.2 and OpenAI's GPT-5.5. On throughput the gap is starker still: 500 repositories scanned in 15 minutes against roughly five hours for frontier models, at a cost Cisco puts at about 172 times cheaper than GPT-5.5. The deployment argument may matter more than the benchmark one. A model that fits on a laptop can run inside the perimeter, which means proprietary source code is never sent to a third-party API for analysis — a constraint that has kept a lot of security teams away from LLM-assisted triage entirely. Open weights make that verifiable rather than promised. The obvious caveat is that these figures are the vendor's own, on a task the vendor selected, and independent replication has not yet appeared. But the direction is one worth watching: a purpose-built small model beating general-purpose giants at a narrow job is a considerably cheaper future than the one where every security workflow calls a frontier API.
Antares-1B →DeepSeek's 1.6-trillion-parameter flagship goes GA — and starts charging by the clock
DeepSeek-V4-Pro left preview on 19 July 2026, three months after its first public release. The weights are the same ones published on 24 April: a sparse mixture-of-experts model with 1.6 trillion total parameters, roughly 49 billion of them active per token, a one-million-token context window and an MIT licence that allows anyone to download and self-host it. What general availability changed was the commercial side. DeepSeek introduced peak-time pricing for the first time, doubling rates during Beijing business hours; off-peak the API costs 0.435 US dollars per million input tokens and 0.87 per million output. Cached input is charged at roughly a hundredth of the off-peak input rate. On 24 July the company retired the legacy deepseek-chat and deepseek-reasoner endpoints, so integrations now have to name a model explicitly. On DeepSeek's own GA figures the model scores 80.6% on SWE-bench Verified — the strongest published result for an open-weights model — with a Codeforces rating of 3,206, 57.9% on SimpleQA Verified and 37.7% on Humanity's Last Exam. Independent measurement tells a subtler story: Artificial Analysis rated V4-Pro at 44 on its Intelligence Index at GA, down from 52 at preview. The model did not get worse; the field moved during the three months it spent in preview. The smaller DeepSeek-V4-Flash and the earlier DeepSeek-R1 remain available, and all three now have profiles in the wujec.ai catalogue.
DeepSeek-V4-Pro →