SmolLM Instruct v0.2 - 135M, 360M and 1.7B parameter instruction tuned Small LMs, Apache 2.0 licensed. Closing the gap to bring intelligence closer to thought (<500 ms per generation)! 🔥

The models are optimised to run on-device with WebGPU support (from MLC and ONNX Runtime) and llama.cpp.

Run them on your Mac, browser, GPU, CPU - it works blazingly fast.

We provide already converted/ quantised - GGUFs, MLC and ONNX checkpoints 🐐

What's new?

We train SmolLM base models on a new synthetic dataset of 2,000 simple everyday conversations we generated by llama3.1-70B -> everyday-conversations-llama3.1-2k

and existing datasets like Magpie-Pro-300K-Filtered by
@argilla_io
, self-oss-instruct-sc2-exec-filter-50k, and a small subset of OpenHermes-2.5 from
@NousResearch


Bonus: We release the fine-tuning scripts we used to train these checkpoints, so that you can fine-tune them for your own use-cases too. ⚡️

Enjoy! and looking forward to what you build with these https://huggingface.co/collections/HuggingFaceTB/local-smollms-66c0f3b2a15b4eed7fb198d0

huggingface.co

💻 Local SmolLMs - a HuggingFaceTB Collection

SmolLM models in MLC, ONNX and GGUF format for local applications + in-browser demos