Hugging Face Releases Full SmolLM3 Training Stack and Checkpoints

Hugging Face has published the complete training stack for SmolLM3, a 3 billion parameter language model designed for resource-efficient inference. The open repository includes scripts for pretraining using Nanotron, post-training alignment with SFT and Adaptive Preference Optimization, and a full suite of evaluation tools. Also included are over 100 training checkpoints, allowing researchers to audit or resume training at any point in the model’s lifecycle.

This release positions SmolLM3 as a viable alternative in the small-to-medium LLM space. Unlike proprietary models, SmolLM3 is fully open and optimized for multilingual (6 languages) and long-context scenarios, with context lengths up to 128,000 tokens. The transparency of its training pipeline makes it particularly valuable for researchers and engineers building or fine-tuning models for on-device or low-latency environments. Hugging Face’s decision to make every layer of SmolLM3 visible—from data to training to alignment—signals a deep commitment to reproducibility in the open-source AI ecosystem.