Granite 4.0 Small, Tiny and Micro Released

• IBM launched Granite 4.0 on 2 October 2025 under Apache 2.0, with Base and Instruct variants of three sizes: Granite-4.0-H-Small, a hybrid mixture-of-experts model with 32B total and 9B active parameters; H-Tiny, 7B total and 1B active; and H-Micro, a 3B dense hybrid^[1]
• Granite-4.0-Micro, a 3B dense model with a conventional transformer architecture, is included for platforms that do not yet support the hybrid design^[1]
• The hybrid models interleave Mamba-2 and transformer layers in a 9:1 ratio, and IBM says this gives over 70% lower RAM use on long inputs and multiple concurrent sessions; the 'H' models use no positional encoding^[1][2]
• IBM positions it as a cost play for agentic work: on Stanford HELM's IFEval, IBM says Granite-4.0-H-Small beats every open-weight model except the 402B-parameter Llama 4 Maverick, and it keeps pace with much larger models on the Berkeley Function Calling Leaderboard v3; vendor-reported results^[1]
• IBM says Granite became the first open model family with ISO/IEC 42001 certification, signs every 4.0 checkpoint on Hugging Face cryptographically and runs a HackerOne bug bounty paying up to $100,000^[1][2]

Notable features

• Compatible with AMD Instinct MI-300X GPUs, and with Qualcomm Hexagon NPUs through work with Qualcomm Technologies and Nexa AI^[1]
• Tiny and Small are IBM's first mixture-of-experts models with shared experts that are always active^[1]
• Early access testers included EY and Lockheed Martin; IBM planned larger and smaller sizes and reasoning variants by the end of 2025^[1]
• Available on watsonx.ai, Hugging Face, Docker Hub, Kaggle, LM Studio, NVIDIA NIM, Ollama, Replicate and Dell platforms, with SageMaker JumpStart and Azure AI Foundry listed as coming^[1]