← All models

Granite 4.0 H Tiny

IBM · 2025-10-02 · 6.9B parameters

Granite 4.0 H Tiny is IBM's instruction-tuned hybrid Mamba-2/Transformer mixture-of-experts model, released on October 2, 2025 and fine-tuned from Granite-4.0-H-Tiny-Base with supervised fine-tuning, reinforcement-learning alignment, and model merging. Its 6,939,037,248-parameter checkpoint combines 36 Mamba-2 layers and four grouped-query-attention layers with 64 routed experts (six active per token) plus a shared expert; IBM describes it as 7B total and 1B active, with a 131,072-token (128K) context. IBM positions it for low-latency, edge, and local applications including tool calling and RAG, and releases the weights under Apache 2.0.

Benchmark scores