Understand this story
On November 3, 2026, Hugging Face proposed the Transformers Model Integration Contract (MIC), packaging three parts—a scaled-down fixture model, runnable Markdown usage examples, and human-reviewed baseline outputs—into model tests, running fast integration tests on CPU. The team used the frozen MIC to replay five months of Transformers commits: during the period when UMT5's fixture regression was introduced on July 30, 2026 and fixed on September 28, it consistently reported a log-probability deviation of 1.22 and a training loss deviation of about 0.20. Llama 4 Scout's scaling report shows that 10.86 billion parameters were scaled down to 44.7 million, with the vision stack reduced from 34 layers to 2 layers and the language stack reduced from 48 layers to 4 layers. As of November 1, 2026, this CI job only reports failures and does not block PRs.
AI-generated from public reports
Latest development · Report
Hugging Face proposes the Transformers Model Integration Contract (MIC)Developments
1 confirmed developments
Hugging Face proposes model integration contract standard transformers-mic
Source:Hugging Face
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 3
Hugging FaceFirst-party
Hugging Face proposes the Transformers Model Integration Contract (MIC)
Hugging Face proposes the Transformers Model Integration Contract (MIC), which packages model tests into three parts—a scaled-down fixture model, runnable Markdown usage examples, and manually reviewed baseline outputs—and runs fast integration tests on CPU. The team used the frozen MIC to replay five months of Transformers commits; the UMT5 fixture consistently reported a log-probability deviation of 1.22 and a training loss deviation of about 0.20 from when the regression was introduced on July 30, 2026 until it was fixed on September 28. The scaled-down report for Llama 4 Scout shows that 10.86 billion parameters were reduced to 44.7 million, the vision stack was reduced from 34 layers to 2 layers, and the language stack was reduced from 48 layers to 4 layers. As of November 1, 2026, this CI job only reports failures and does not block PRs.