SmolLM3
Open 3B parameter language model with reasoning modes
SmolLM3 is a 3 billion parameter decoder-only transformer released by Hugging Face. It supports six languages and long context windows up to 128k tokens. The model offers dual reasoning modes and is fully open with public training details.
- 3 billion parameters with open weights and training details
- Supports context windows up to 128k tokens via YaRN
- Natively supports English, French, Spanish, German, Italian, Portuguese
- Dual mode reasoning with optional extended thinking traces
- Trained on 11.2 trillion tokens including code and math data
What it is
SmolLM3 is a language model with 3 billion parameters developed by Hugging Face. It is designed to perform well at a small scale while supporting advanced features like long context processing and hybrid reasoning. The model handles text in six European languages and can interact with external tools. It is fully open, meaning the weights and training configurations are publicly available for anyone to use or study.
Inside the model
The architecture is a decoder-only transformer using Grouped Query Attention and No Positional Embeddings with a three-to-one ratio. Pretraining involved 11.2 trillion tokens from web, code, math, and reasoning sources in a staged curriculum. Post-training included midtraining on 140 billion reasoning tokens, followed by supervised fine-tuning. Alignment used Anchored Preference Optimization. The model supports context lengths up to 128k tokens via YaRN extrapolation, though it was trained on 64k.
What it can do
SmolLM3 operates in two modes: standard and extended thinking. In reasoning mode, it scores 36.7 on AIME 2025 and 83.4 on GSM-Plus. Without extended thinking, it achieves 76.7 on IFEval and 92.3 on BFCL for tool calling. It supports agentic usage through XML or Python-style tool calls. Benchmarks show strong performance in instruction following and multilingual question answering compared to similar sized models.
How to use it
The model weights are fully open and available on Hugging Face. Users can run it locally with libraries like llama.cpp, ONNX, MLX, or ExecuTorch. It is compatible with transformers version 4.53.0 and later. Deployment options include vLLM and SGLang for API-compatible serving. Quantized checkpoints are provided for efficient local inference on various hardware setups.
Questions people ask about SmolLM3
What languages does SmolLM3 support?
SmolLM3 natively supports six languages: English, French, Spanish, German, Italian, and Portuguese. It is designed for multilingual tasks within these European languages.
How do I enable reasoning mode in SmolLM3?
Extended thinking is enabled by default. You can control it using the /think or /no_think flags in the system prompt, or by setting the enable_thinking keyword argument when generating responses.
What is the maximum context length?
The model was trained on a context length of 64k tokens but supports up to 128k tokens using YaRN extrapolation. Configuration files allow adjusting this for even longer inputs if needed.
Good reading and sources
There is no English Wikipedia article on SmolLM3 yet, so the facts on this page come from the sources listed above: huggingface.co/HuggingFaceTB/SmolLM3-3B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.