IBM Granite 4.1 3B

Granite-4.1-3B is the entry size of IBM's Granite 4.1 family, published on 29 April 2026. It keeps the property that matters most for small deployments: the full 131,072-token context of its larger siblings, in a model of 3.4 billion parameters that runs on a laptop. The shape differs slightly from the bigger models. Embedding size drops to 2560 and the model uses 40 attention heads rather than 32, still with 8 key-value heads and 40 layers. Vocabulary, position scheme and licence are identical to the rest of the family. IBM's evaluation shows the cost of the smaller size honestly: 67.02 on MMLU against the 8B's 73.84, 31.70 on GPQA against 41.96, and 60.80 on BFCL v3 tool calling against 68.27. Code generation holds up better than knowledge — 81.71 on HumanEval, within four points of the 8B — which is the usual pattern for models this size. Multilingual MMLU at 57.61 is the weakest result of the three and the one to watch if the deployment is not in English. Its role in the catalogue is as the reference point for what a 2026 small model gives you under a licence with no strings: no territorial exclusion, no revenue threshold, no acceptable-use annex. It was superseded by Granite 4.2 3B in August 2026, which added switchable thinking modes; IBM has not withdrawn this version.

▸Videos

No videos yet.

▸Reviews