MDL-9101EST.2026 · IDX.154
Language modelIn production

IBM Granite 4.2 30B

IBM · USA · 2026

IBM's largest openly published reasoning model: 30 billion dense parameters under a plain Apache 2.0 licence, with a thinking mode that can be dialled down or switched off per query.

wujec.ai score

/10

Community score

no votes yet
Sign in to rate

Granite-4.2-30B is the flagship of the Granite 4.2 family, published by IBM's Granite Team on 25 August 2026. It is the first generation of the family with reasoning built into the model itself: it writes out a chain of thought inside <think> tags before answering, and the user can choose between full thinking, a low-effort mode and no thinking at all on a per-query basis. Where most vendors ship a separate reasoning model, IBM ships one model with a dial. The architecture is deliberately conservative. At a time when almost every model of this size is a mixture of experts that activates a fraction of its weights per token, Granite 4.2 stays dense: all 29.28 billion parameters work on every token, across 64 layers with grouped query attention. That costs more compute per token, but it removes the routing machinery that makes mixture-of-experts models awkward to serve on modest hardware and unpredictable under load. IBM declares 128K tokens of native context with an extension to 512K, and reports 81.38 on RULER at the full 128K, which is a genuinely strong long-document result for this size. On IBM's own evaluation runs the 30B resolves 57 percent of SWE-bench Verified and 33.29 of the harder SWE-bench Pro, scores 89.17 on AIME25 and 66.41 on GPQA. These are the manufacturer's numbers, run on an NVIDIA-derived evaluation framework, and we report them as such. Training itself was done on rented hardware: an NVIDIA GB200 NVL72 cluster at CoreWeave, which is worth noting for a company that sells its own mainframes. For a reader in Poland the licence is the headline. Granite carries plain Apache 2.0 with no territorial carve-outs, no user thresholds and no separate acceptable-use policy - the repository does not even contain a licence file with extra clauses, only the standard declaration. That is a meaningfully different proposition from the open-weight models whose agreements exclude the European Union outright. The catch is elsewhere: IBM lists twelve tested languages - English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese - and Polish is not one of them. The model may well handle Polish; the manufacturer simply does not claim it does.

#open-weights#reasoning#enterprise#apache-2.0#IBM
Official website

Videos

No videos yet.

Reviews

No reviews yet. Be the first!

Sign in to write a review