From the source
Hugging Face releases SmolLM3, a fully open 3B small language model that outperforms Llama-3.2-3B and Qwen2.5-3B while being competitive with larger 4B models.
It supports multilingual (6 languages), long context up to 128k, and dual mode reasoning (think/no_think).
The release includes a complete engineering blueprint with architecture details, data mixtures, and training methodology.
