IBM Granite 4.0 H Micro

Granite-4.0-H-Micro is the smallest of the three hybrid models IBM published in October 2025 and the one with the longest commercial afterlife. It holds 3.19 billion parameters, all of them active — there is no mixture of experts here, only the hybrid layer stack: four attention layers and 36 Mamba-2 layers, embedding size 2048, 32 attention heads with 8 key-value heads, and no positional encoding. Two properties make it useful at this size. The first is the fixed-size recurrent state, which keeps the memory cost of a 128K-token input flat rather than letting it grow with every token; on a laptop or an edge device that is the difference between running and not running. The second is that IBM did not cut the language coverage: the same twelve tested languages as the flagship, and MMMLU 55.19 against the flagship's 69.69 — lower, but present. Its benchmark profile is that of a small model that punches at its weight: MMLU 67.43, MMLU-Pro 43.48, GPQA 32.15, IFEval average 84.32 (better than the seven-billion H Tiny), ArenaHard 36.15, HumanEval 81, MBPP 73, BFCL v3 57.56 for tool calling. On instruction following it is the second-best of the whole generation. It is also the only Granite 4.0 model still listed by the large API resellers, at roughly 0.017 dollars per million input tokens and 0.112 per million output — cheap even by the standards of small models, and cheaper than IBM's own newer 4.1 and 4.2 offerings. For a reader who wants to try the Mamba-2 hybrid without hosting anything, this is the only door of this generation still open. The licence is Apache 2.0 with no added conditions. Polish is not among the tested languages.

▸Videos

No videos yet.

▸Reviews