Tier 4 Deeply open
What the publisher claims: fully open, open-source
The one-liner: Small model with an unusually complete public engineering blueprint.
The six questions
weights downloadable, ungated
Apache 2.0
standard Transformers, vLLM and ONNX paths
recipe, data mixtures, training framework and alignment method published, but the full training corpus and checkpoints are not
a skilled group could closely approximate, not reproduce exactly
Hugging Face ships SmolLM with its whole training stack public — nanotron, datatrove, and lighteval all carry contribution guides and merge outside pull requests — and solicits community input on model direction, though release decisions remain Hugging Face's.
License
Apache 2.0 OSI-approved — license source ↗
What is actually released
- base and instruct weights
- blog whiteprint with architecture and data mixtures
- public datasets used (FineWeb-Edu and others)
- nanotron and datatrove training framework
- model card describing midtraining
- SFT and Anchored Preference Optimization alignment
Why it matters
Demonstrates that meaningful reproducibility at small scale is achievable without publishing every pretraining byte.
Sources
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.