IBM Granite 4.0 Nano releases sub-billion edge LLMs with Apache 2.0 and runtime support

AI Impact Summary

IBM's Granite 4.0 Nano introduces the smallest members of the Granite 4.0 family, targeting on-device and edge runtimes. The Nano lineup includes Granite 4.0 H 1B (~1.5B parameters) and Granite 4.0 H 350M (~350M parameters), plus traditional transformer equivalents Granite 4.0 1B and Granite 4.0 350M, all released under Apache 2.0 and supported on vLLM, llama.cpp, and MLX. Trained with the same methodologies and 15T tokens, they leverage Granite's hybrid-SSM architecture and come with ISO 42001 certification for governance. Benchmarks claim improved performance at small scales across knowledge, math, code, and safety, and strong results on instruction following and tool calling compared to similarly sized models, indicating readiness for edge deployment where model size was previously a constraint.