●Change8

v4.49.0-Mistral-3

📦 transformersView on GitHub →
✨ 7 features🔧 4 symbols

Summary

This release introduces Mistral 3 (Mistral Small 3.1) to the Transformers library, a 24B parameter model featuring 128k context length and advanced vision-language capabilities.

Migration Steps

  1. Install the specific version from the git tag: pip install git+https://github.com/huggingface/transformers@v4.49.0-Mistral-3

✨ New Features

  • Added support for Mistral 3 (Mistral Small 3.1) model architecture.
  • Enhanced long context capabilities up to 128k tokens.
  • Added state-of-the-art vision understanding capabilities.
  • Support for image-text-to-text pipeline inference.
  • Support for batched image and text inputs, including varying numbers of images per text.
  • Support for 4-bit quantization via BitsAndBytesConfig.
  • Support for chat templates with multi-modal inputs.

🔧 Affected Symbols

AutoModelForImageTextToTextAutoProcessorpipelineMistral3