Granite 4.0 Nano: Just how small can you go?

Source: HuggingFace Date: 2025-10-28 URL: https://huggingface.co/blog/ibm-granite/granite-4-nano

Summary

Model release from IBM: Granite 4.0 Nano family, four variants at 350M and ~1.5B parameters in both hybrid-SSM (Granite 4.0 H) and traditional Transformer architectures. Trained on 15T tokens. Claims outperform Qwen, LFM, and Gemma at equivalent sizes on IFEval and BFCLv3 (function calling) benchmarks. Apache 2.0, supported in vLLM, llama.cpp, MLX. ISO 42001 certified.

Implications

Thread: open-weights ecosystem health / model release cadence. IBM continues pushing Granite as the enterprise-safety-certified alternative in the small model space. The hybrid-SSM variants (H 1B, H 350M) are the architecturally interesting models — hybrid attention+SSM at 350M is a meaningful architecture test at the smallest scale anyone has attempted it. The BFCLv3 function calling emphasis signals IBM is targeting agentic edge deployment, not just chat. ISO 42001 certification is a differentiator for regulated industry procurement. Specific benchmark numbers not published in the post is a gap.