Granite 4.0 Nano: Just how small can you go? (huggingface.co)

IBM released Granite 4.0 Nano, the smallest models in its Granite 4.0 family aimed at edge and on-device use. The suite includes four instruct models and base counterparts: hybrid-SSM variants Granite 4.0 H 1B (~1.5B params) and H 350M (~350M params), plus traditional transformer 1B and 350M versions for runtimes that lack hybrid support (e.g., llama.cpp). All Nano models inherit Granite 4.0’s improved training pipeline and over 15 trillion token training mix, are Apache 2.0–licensed, certified under ISO 42001 for responsible model development, and ship with native runtime support (vLLM, llama.cpp, MLX). Full model cards are available on Hugging Face. Technically notable is the hybrid-SSM architecture: it delivers strong capability-per-parameter, which IBM shows across benchmarks in general knowledge, math, code and safety, and outperforms similarly sized competitors (e.g., Qwen, LFM, Gemma) on agentic tasks like instruction following and tool calling (IFEval, BFCLv3). The inclusion of transformer variants ensures broad runtime compatibility while the hybrid models push efficiency for constrained environments—making Granite 4.0 Nano significant for developers building local, low-latency, privacy-sensitive, or resource-limited AI applications.