Recently, the IBM AI team launched the Granite4.0Nano series, a family of small models designed for local and edge inference, aiming to provide enterprises with stronger control and open-source licensing. This series includes 8 models, available in two sizes: 350M and approximately 1B, using a hybrid SSM and transformer architecture, supporting both basic and instruction modes. All models are released under the Apache2.0 license and can run natively in popular runtime environments such as vLLM, llama.cpp, and MLX.

The Granite4.0Nano series includes four model lines and their base versions. Among them, Granite4.0H1B uses a hybrid SSM architecture with approximately 1.5B parameters, while Granite4.0H350M also uses a hybrid approach with 350M parameters. To ensure maximum runtime compatibility, IBM also provides transformer versions of Granite4.01B and Granite4.0350M.

These H variant models alternate between SSM layers and transformer layers, and this hybrid structure has significant advantages in memory growth compared to pure transformers, while maintaining the generality of the transformer modules. The Granite4.0Nano models did not use reduced data pipelines, but followed the same training method as the large-scale Granite4.0 models, trained on over 15 trillion tokens, and underwent instruction tuning to enhance tool usage and instruction following capabilities.

image.png

IBM also compared the Granite4.0Nano with other similar models, including Qwen, Gemma, and LiquidAI LFM. The results showed significant improvements in areas such as general knowledge, mathematics, code, and security. Additionally, the series of models performed well in agent tasks on the IFEval and Berkeley function call leaderboard version 3.

These models are ISO42001 certified and released with cryptographic signatures, ensuring traceability and governance capabilities required for enterprise use. Users can access these models through Hugging Face and IBM watsonx.ai and deploy them at the edge, local, and browser levels, helping early AI engineers and software teams better implement projects.

huggingface:https://huggingface.co/collections/ibm-granite/granite-40-nano-language-models

Key Points:  

🔹 IBM released the Granite4.0Nano series, containing 8 models suitable for edge AI inference.  

🔹 The series of models were trained on over 15 trillion tokens, ensuring performance inheritance.  

🔹 All models are ISO42001 certified and have enterprise-level governance capabilities, supporting multiple runtime environments.