France · Hugging Face
Transformers-mic is a proposed
Compiled by KHAO Editorial — aggregated from 1 source. See llms.txt for citation guidance.
★ Tier-1 Source
The contract holds if running the usage path on the fixture still produces the baselines.
Key facts
- The Llama 4 Scout shrink report shows what this looks like on a big model. 108.6 billion parameters become 44.7 million
- In their CPU study, the experimental UMT5 fixture used 66 MB of weights; the source checkpoint is 1,227 MB
- The small UMT5 fixture flagged the regression from its introduction on July 30, 2026 until its fix on September 28
- On July 30, 2026, Transformers enabled PyTorch's scaled dot-product attention (SDPA) for T5
Summary
On July 30, 2026, Transformers enabled PyTorch's scaled dot-product attention (SDPA) for T5. When they later replayed those commits, a small UMT5 model running on CPU flagged it on every one of them. Transformers supports over 750 model types, and its test suite is big: 88,000+ CPU tests on main, plus. The integration tests load real checkpoints and exercise the model's API on real hardware. Their approach is to shrink the model instead. In their CPU study, the experimental UMT5 fixture used 66 MB of weights; the source checkpoint is 1,227 MB.