Granite 4.0 Tiny

Granite 4.0 Tiny (Granite-4.0-H-Tiny) is IBM's efficiency play in the Granite 4.0 family, launched October 2, 2025. It is a 7B-total, ~1B-active hybrid mixture-of-experts model that interleaves Mamba-2 state-space layers with a minority of transformer attention blocks (9:1 ratio), removing positional encodings entirely for better length generalization.

IBM claims the hybrid design cuts memory for long-context and multi-session inference by over 70% versus conventional transformers, and the family is the first open LLM line to receive ISO 42001 certification, with cryptographically signed checkpoints and a HackerOne bug bounty. The Tiny variant carries a 128K context window under Apache 2.0.

Honest limits: memory-reduction claims are IBM-reported and workload-dependent. The Mamba-hybrid stack is less universally supported by inference tooling than pure transformers, and ecosystem benchmarks trail the major open families.