Hugging Face releases SmolLM3
The 3-billion-parameter model lets users toggle reasoning on or off per query and scored 36.7% on AIME 2025 with reasoning enabled versus 9.3% without.
- Open weights & ecosystem
- Models & capabilities
- Minor
Hugging Face released SmolLM3, a 3-billion-parameter open-weight language model published as base and instruction-tuned variants. It extended the company’s SmolLM line of small, efficient models with two features more commonly found in much larger systems: a reasoning-mode switch, toggled with /think and /no_think flags so a user can choose between extended chain-of-thought and a direct answer; and a context window of up to 128,000 tokens, achieved with techniques including NoPE and YaRN rather than a proportionally larger model.
Hugging Face reported that the reasoning mode mattered substantially on harder tasks: on the AIME 2025 maths competition benchmark, SmolLM3 scored 36.7% with reasoning enabled against 9.3% without it. The company said the model outperformed similarly sized rivals including Llama-3.2-3B and Qwen2.5-3B, and was competitive with some 4-billion-parameter alternatives, while adding support for six languages: English, French, Spanish, German, Italian and Portuguese.
The release continued a trend through 2025 of open small models absorbing capabilities — extended reasoning, long context, multilingual support — that had first appeared only in frontier closed models, narrowing the practical gap between what ran on a single consumer GPU and what required a data centre.