IBM Granite 4.0 H Tiny

Granite-4.0-H-Tiny is the model of this generation with the widest gap between what it stores and what it computes. The published weights hold 6.94 billion parameters, but the mixture of experts routes only six of its 64 experts per token, so roughly one billion parameters do the work on any given word. In practice that means a model which needs the memory of a 7B model and the compute of a 1B one. Like the rest of the hybrid line it uses four attention layers and 36 Mamba-2 state-space layers, with no positional encoding. Embedding size is 1536, the attention layers carry 12 heads with 4 key-value heads, and the sequence length is 128K tokens. Because the Mamba-2 state is fixed in size, the memory cost of a long document does not climb the way it does in an ordinary transformer — the property that matters most on the consumer hardware this size is aimed at. The scores sit where the size suggests, and occasionally above it: MMLU 68.65, MMLU-Pro 44.94, GPQA 32.59, IFEval average 81.44, HumanEval 83, MBPP 80, BigCodeBench 41.06. Two results are worth singling out. On multilingual grade-school maths (MGSM 45.36) it beats the 32-billion-parameter flagship of its own family, and on the multilingual code benchmark MULTIPLE it scores 55.83, close to the flagship's 57.37. It is also the generation's most downloaded hybrid: about 90,800 pulls of the main repository in the thirty days to 1 September 2026, well ahead of the flagship's 25,700. For a model released in October 2025 and superseded twice since, that is a strong sign it found a niche — small enough to self-host, capable enough to be useful, and free of licence conditions. The licence is plain Apache 2.0. Polish is not among the twelve languages IBM tested.

▸Videos

No videos yet.

▸Reviews