SmolLM

Published by Hugging Face under its HuggingFaceTB organization, this project bundles a family of compact, fully open language models with the code used to train them. The current flagship, SmolLM3, has 3 billion parameters, saw roughly 11 trillion training tokens, and ships as a base model plus an instruct model that toggles between an extended thinking mode and a direct answer mode. It covers English, French, German, Spanish, Italian, and Portuguese and reaches 128k tokens of context via YaRN extrapolation. The earlier SmolLM2 generation offers 135M, 360M, and 1.7B sizes small enough for phones, browsers, and CPU-only machines. The GitHub repo collects pretraining and post-training configs, evaluation scripts, data curation notes, and local inference examples for llama.cpp, MLX, MLC, and Transformers.js, and it also hosts the SmolVLM vision-language models. Weights are on the Hugging Face Hub under Apache 2.0, which allows commercial deployment, and the pretraining and instruction datasets are public as well. Choose it for on-device assistants, edge prototypes, or as a worked reference for how small models are built end to end.