SmolLM2 is a family of compact language models available in three sizes, capable of a wide range of tasks while being lightweight for on-device use. The most powerful model is SmolLM2-1.7B-Instruct, which can be used as an assistant with transformers, trl, or quantized versions with tools like llama.cpp, MLX, and transformers.js. SmolLM2-360M and SmolLM2-135M are lighter models suitable for on-device usage. The collection includes transformers, local applications, pre-training, fine-tuning, and evaluation resources. SmolTalk, the SFT dataset, and synthetic data pipelines are highlights of the content, showcasing unique and valuable contributions to the AI research field.
https://github.com/huggingface/smollm