Model
Inference & Serving
SmolLM2
1Hugging 2Face's 3compact 4on-device 5open 6model
the six Ws · specification
W1
Who
Hugging Face's HuggingFaceTB team developed the SmolLM2 model family.
W2
What
SmolLM2 is a family of small LLMs (135M, 360M, and 1.7B parameters) trained on 11 trillion tokens for on-device use.
W3
Where
Weights are on Hugging Face under the HuggingFaceTB organization, with training and eval code at github.com/huggingface/smollm.
W4
When
Released in November 2024.
W5
Why
Designed to run capable language models locally on phones and laptops without cloud inference.
W6
With
Runs via Hugging Face Transformers, ONNX, or llama.cpp/GGUF on CPU or mobile hardware.
W7
Watch
Apache 2.0 license, fully open weights and training data, no credentials required, runs locally; maintained by Hugging Face.
alternatives
works with