Granite 4.0

Granite 4.0, open-sourced October 2, 2025, is IBM's first departure from pure-Transformer design: a hybrid architecture interleaving Mamba-2 state-space layers with transformer attention blocks in a 9:1 ratio. The family ships four models — Granite-4.0-Micro (3B dense transformer), H-Micro (3B hybrid dense), H-Tiny (7B total, ~1B active MoE), and H-Small (32B total, ~9B active MoE) — plus later Nano variants at 350M/1B for on-device and in-browser inference.

IBM reports over 70% less RAM for long-context and multi-session inference versus conventional transformers. The enterprise governance story is the real differentiator: Apache 2.0 license, ISO 42001 certification (claimed first for an open model family), cryptographically signed checkpoints, and a HackerOne bug bounty up to $100,000. All sizes support 128K context and strong function-calling for agentic enterprise workloads.