SmolLM3

SmolLM3 is a compact language model from Hugging Face for developers building local AI applications and researchers studying model training. It has 3B parameters and comes in base and instruction-tuned variants.

• Direct or reasoning responses: The instruction model can answer directly or produce an explicit reasoning trace, with control over the mode during a conversation.
• Long context: A context window of up to 128k tokens provides room for lengthy documents, code and conversation history.
• Multilingual text: The model supports English, French, Spanish, German, Italian and Portuguese.
• Tool calling: Applications can give the instruction model tool definitions so it can request calls to those tools.
• Training transparency: Hugging Face publishes the model weights, architecture details, public training data mixture and training configurations for researchers who want to examine or build on the model.

SmolLM3 belongs to Hugging Face's family of models designed for on-device use. The project repository uses the Apache 2.0 license. Its maximum context window uses YaRN to extend beyond the 64k-token context length used in training.