IBM Granite 4.0

Open-sourced October 2, 2025, Granite 4.0 is IBM's first departure from pure transformers: a hybrid architecture interleaving Mamba-2 state-space layers with a minority of attention layers in a 9:1 ratio, with no positional encodings. The family spans Nano (350M/1B, browser-capable) to H-Small (32B total, 9B active MoE), all under Apache 2.0. IBM claims over 70% less memory for long-context and multi-session inference versus conventional transformers, trained on 22T tokens with context validated to 128K. Notably, Granite is the first open model family to receive ISO 42001 certification, with cryptographically signed checkpoints and a HackerOne bug bounty up to $100,000.\n\nLimitations: it is enterprise-positioned rather than benchmark-leading — it won't top reasoning leaderboards against frontier models. The 9:1 hybrid is a newer recipe with less community fine-tuning support than Llama/Qwen. Some sizes launched in public beta, with Medium and Thinking variants arriving later in 2025.

IBM-reported: 70%+ memory reduction vs transformers; Nano H-1B 78.5 IFEval