SmolLM3
A 3B model with its blueprints published
What it is
Hugging Face's fully open 3-billion-parameter model with a toggleable extended-thinking mode, 64k context natively (128k+ via YaRN), six languages, and tool calling — trained on 11T tokens with the complete recipe, configs, and data mix published.
Why it's interesting
Arguably the best-documented small model in existence: not just open weights but open engineering — the nanotron, datatrove, and TRL recipes are all public, making it a reference for anyone studying modern LLM training at accessible scale.
Use cases
- On-device assistants and agents with tool calling
- A fine-tuning base for domain-specific small models
- Studying a full modern training pipeline end to end
Who it's for
Edge and on-device developers, researchers, fine-tuners
Setup
Easy. Runs in transformers, llama.cpp, vLLM, or Ollama; ~6GB in bf16, ~3GB quantized
Limitations & cautions
A 3B-class reasoning ceiling, strongest in its six core languages, and long context beyond 64k depends on extrapolation quality.
Editorial takeaway
Small models are where open source wins outright — and this one ships with its homework attached.