Hugging Face releases small language model with full training recipe

Hugging Face has launched SmolLM3, a new 3-billion-parameter language model designed for high performance and efficiency. In their official post, the company states that the model outperforms others in its size class and is competitive with some larger alternatives. A key feature is its dual-mode capability, which allows it to provide direct answers or show its reasoning process. SmolLM3 also supports long texts up to 128,000 tokens and understands six European languages. Unusually, Hugging Face also published the complete “engineering blueprint”, detailing the exact data and methods used to build the model, to encourage further research and development by the community.