Hugging Face releases SmolLM3, a small reasoning model

Background

SmolLM was a Hugging Face series of small models designed for devices. Its second generation, SmolLM2, came before it. SmolLM3 extended the line with a reasoning mode that users can switch on or off ^[1].

What happened

On 8 July 2025 Hugging Face released SmolLM3, a 3-billion-parameter open model ^[1]. A parameter is an adjustable number inside a model that stores part of what it has learned, so 3B means about three billion such numbers. The model was trained on about 11 trillion tokens and supports six European languages: English, French, Spanish, German, Italian and Portuguese ^[1].

The release was framed as a model that is small, multilingual and able to handle long inputs. Those three features distinguish it from many models of similar size.

Key details

SmolLM3 can handle up to 128K tokens of context, using YaRN extrapolation to reach that length ^[1]. A context window is the amount of text a model can consider at one time, and 128K tokens is enough for a long report or a large set of code files. A token is a small unit of text, usually a word or part of one.

Users can choose between a reasoning mode and a direct-answer mode with the /think and /no_think flags ^[1]. Reasoning mode lets the model work through a problem before answering, which usually takes more time but can improve results on hard questions. The direct mode gives a quicker reply.

Hugging Face said that SmolLM3 beats Llama-3.2-3B and Qwen2.5-3B and competes with 4B models ^[1]. These comparisons are the developer's own, so readers should check how they were measured before drawing firm conclusions.

Impact

The full training recipe, the data mixtures and the checkpoints were released with the model ^[1]. Publishing the process rather than only the final weights means other researchers can examine how the model was made, repeat parts of the work and compare methods. For people who want to learn how small models are built, the release is a practical reference.

The event also showed that a 3B model could combine several languages, long context and an optional reasoning mode in a single open release. For a reader following the field, SmolLM3 is a clear point in the path from the earlier SmolLM models to the more capable ones that followed.