IBM Releases Granite 4.0 Nano

IBM has released Granite 4.0 Nano, the smallest models in the Granite 4.0 model family. Optimized for edge devices and on-device applications, these models are provided under the Apache 2.0 license.

The key model lineup is as follows:

• Granite 4.0 H 1B & 350M: Models adopting a hybrid-SSM (hybrid-SSM) based architecture to maximize efficiency.
• Granite 4.0 1B & 350M: Versions using the conventional Transformer (Transformer) architecture, provided for environments with inadequate hybrid architecture support (e.g., llama.cpp).

In terms of performance, Granite 4.0 Nano has demonstrated superior performance compared to similarly sized models such as Qwen and Gemma on general knowledge, math, code, and safety benchmarks. In particular, it showed excellent capabilities essential for agentic workflows on the IFEval (instruction following) and BFCLv3 (tool calling) benchmarks.

Additionally, it complies with responsible model development standards through ISO 42001 certification, and provides native architecture support across major runtimes including vLLM, llama.cpp, and MLX.