IBM Granite 4.0 Micro

Granite-4.0-Micro is the odd one out of its own generation, and that is the point of it. While H Small, H Tiny and H Micro replaced 36 of their 40 layers with Mamba-2 state-space blocks, Micro keeps all 40 as ordinary attention layers with RoPE positions — a conventional dense transformer of 3.40 billion parameters, embedding size 2560, 40 attention heads with 8 key-value heads. IBM shipped it for the large amount of inference software that had no support for hybrid architectures in late 2025. The cost of that choice is visible in exactly the place theory predicts. Micro matches its hybrid twin on general knowledge (MMLU 65.98 against 67.43) and beats it on reasoning chains (BBH 72.48 against 69.36) and on maths (GSM8K 85.45 against 81.35), but it pays for a 128K-token context with a key-value cache that grows with every token, where the hybrid keeps a state of fixed size. On multilingual grade-school maths the gap runs the other way and is large: MGSM 28.56 against 44.48. It is the most downloaded and most bookmarked model of the generation — roughly 59,500 pulls of the main repository in the thirty days to 1 September 2026 against 19,100 for its hybrid twin — which says less about quality than about how much software still expects a plain transformer. There is a longer point here for anyone following the family. Granite 4.1 in April 2026 and Granite 4.2 after it are dense attention-only models with RoPE, and the smallest of them shares this model's layer count, embedding size and head count. The hybrid branch of Granite 4.0 was the experiment; this quiet conventional model was the branch that survived. The licence is Apache 2.0 with no added conditions. Polish is not among the twelve tested languages.

▸Videos

No videos yet.

▸Reviews