IBM on Thursday open-sourced Granite 4.0, a family of language models built on a hybrid architecture that combines Mamba-2 state-space layers with a minority of transformer attention layers in a 9:1 ratio. The release includes four models: Granite-4.0-H-Small (32B total parameters, 9B active), Granite-4.0-H-Tiny (7B total, 1B active), Granite-4.0-H-Micro (3B dense hybrid), and Granite-4.0-Micro (3B dense transformer). All are available under the Apache 2.0 license.

IBM says the hybrid design can offer over 70 percent reduction in RAM needed to handle long inputs and multiple concurrent batches compared to conventional transformer models, and claims the models can run on cheaper GPUs at lower cost. The company also says Granite is the first open model family to receive ISO 42001 certification, following an extensive, months-long external audit of IBM’s AI development process. Model checkpoints on Hugging Face are cryptographically signed, and IBM has partnered with HackerOne on a bug bounty program offering up to $100,000 for identified vulnerabilities.

The models are available on IBM watsonx.ai and through partners including Hugging Face, NVIDIA NIM, Ollama, and Replicate. Support for Amazon SageMaker JumpStart and Microsoft Azure AI Foundry is coming soon. IBM reports that Granite-4.0-H-Small outperforms all open-weight models except Meta’s 402B-parameter Llama 4 Maverick on Stanford HELM’s IFEval instruction-following benchmark, and keeps pace with larger models on the Berkeley Function Calling Leaderboard v3. Early access partners included EY and Lockheed Martin.

IBM says the launch doubles down on small, efficient language models that provide competitive performance at reduced costs and latency.