Model Integration Contracts (MIC)

huggingface-mic hosts the tiny models and contracts that let 🤗 Transformers run fast, meaningful integration tests on every pull request.

For each model architecture, we build a slim fixture, a very small version of a real checkpoint, together with an executable contract (the usage code for the model, plus the outputs it must produce). Baselines are measured on CPU and CUDA and reviewed by a human before they are accepted. They are never fitted to make a test pass.

• Model repos (<name>-mic): the slim fixture, the contract and its reviewed expectations.
• transformers-fleet: the pinned set of contracts that Transformers CI runs.
• mic-coverage: which model types have a contract, which can get one, and how each fixture was shrunk.
• Code and docs: github.com/huggingface/transformers-mic

Add a model

pip install -U "transformers-mic @ git+https://github.com/huggingface/transformers-mic.git@main"
hf auth login     # write token
tmic auth login   # GitHub login
tmic run Qwen/Qwen3-0.6B

tmic run drafts the contract, builds the slim fixture and measures baselines in the cloud, so you don't need a GPU. Before anything is published, it asks you to review the contract and any changed values. Model experts are especially welcome: a shrinking recipe can leave a tiny model useless, like a weather model that forecasts sunshine forever and passes every test.

For now, contributing is restricted to members of the huggingface organization on the Hub.