← All models

SmolLM3-3B

Hugging Face · 2025-07-08 · 3.1B parameters

SmolLM3-3B is Hugging Face's open 3B dual-mode instruct and reasoning model, built on SmolLM3-3B-Base, which was pretrained on 11.2T tokens and then mid-trained for long context and on about 140B reasoning tokens. Post-training combined supervised fine-tuning, Anchored Preference Optimization, and model merging; extended thinking is on by default and can be switched with /think and /no_think system flags or the enable_thinking template argument. The dense decoder uses GQA with RoPE removed from every fourth layer, trains at 65,536 tokens, and is documented to handle 131,072 tokens with 2x YaRN extrapolation.

Benchmark scores