IBM Granite 4.1 30B

Granite-4.1-30B is the top of IBM's Granite 4.1 family, published on 29 April 2026 — four months before the Granite 4.2 generation that replaced it. This catalogue documents it because the 4.1 models are still downloaded in six-figure numbers every month and remain the version many production deployments were built on. The design is deliberately unfashionable: a dense decoder-only transformer, 64 layers, embedding size 4096, 32 attention heads with 8 key-value heads, a 131,072-token sequence length and RoPE positions. Dense means every one of the 28.87 billion parameters runs on every token, with none of the mixture-of-experts routing that its Chinese rivals of the same season used to cut inference costs. What you gain is predictability: memory use and latency do not depend on which experts a prompt happens to wake up. IBM's own evaluation puts it at 80.16 on MMLU, 64.09 on MMLU-Pro, 45.76 on GPQA, 89.65 average on IFEval, 88.41 on HumanEval and 73.68 on BFCL v3 for tool calling. Multilingual MMLU comes in at 73.71 across eleven languages. All of these are the maker's figures. The difference against the 4.2 generation is not size but behaviour. Granite 4.2 added a built-in chain of thought with switchable thinking modes; 4.1 has none — it is an instruction model that answers directly. For workloads where a visible reasoning trace is unnecessary overhead, that is a feature rather than a shortcoming. The licence is plain Apache 2.0: no territorial exclusion, no revenue threshold, no separate acceptable-use document. For a European reader that remains the least encumbered footing on which frontier-adjacent weights are published.

▸Videos

No videos yet.

▸Reviews