SmolLM2 1.7B Instruct

SmolLM2 is Hugging Face's small-model line, designed specifically for on-device and edge deployment. The 1.7B-Instruct sits at the top of the family, with 135M and 360M sibling sizes for even more constrained environments.

Why care about a 1.7B model

Most modern models prioritize capability over efficiency. SmolLM2 inverts that: it's a model you can run at acceptable speed on a phone, a Raspberry Pi 5, or a Jetson Nano. For simple summarization, classification, command parsing, and on-device assistants, that tradeoff is the right one.

What it's good at

Short-form instruction-following, summarization, simple Q&A, basic structured-output generation, and as a teaching aid for understanding small-model behavior. It will not reason or do anything that requires long-horizon planning.

Training story

Trained on 11 trillion tokens — extraordinary for its size class — drawing heavily on the FineWeb-Edu corpus and synthetic educational content. The training mix was published alongside the model and is itself a useful research artifact.

License

Apache 2.0 — fully permissive.