SmolLM3: an open 3B reasoner with the full recipe
Hugging Face released SmolLM3, a 3B model trained on 11T tokens with dual think/no_think modes, six languages and 128K context, published with its complete training blueprint.
What happened
Hugging Face released SmolLM3, a fully open 3B model, with base and instruct/reasoning checkpoints.
The technical details
From the post:
- trained on 11T tokens; reported to beat Llama-3.2-3B and Qwen2.5-3B and stay competitive with 4B models (Qwen3, Gemma3)
- the instruct model has dual-mode reasoning:
think/no_think - 6 languages: English, French, Spanish, German, Italian, Portuguese
- long context up to 128K, using NoPE and YaRN
- the complete recipe: architecture details, exact data mixtures across a three-stage pre-training approach, and how the hybrid reasoning model was built
Why it matters
It’s the successor to SmolLM2 (summarised here in Feb 2025), and it follows Qwen3’s switchable thinking mode (May 2025) at a much smaller size. The published recipe makes it a useful reference for anyone training small models.
Sources
- SmolLM3: smol, multilingual, long-context reasoner — Hugging Face
This is a summary written from the linked public sources. Figures and numbers are as reported there.