HuggingFace × .txt: Structured generation in Rust with outlines-core

Structured generation is now faster and lighter with outlines-core.

Structured generation has quickly become a foundation for building reliable LLM applications. As adoption of Outlines grew, we gathered user feedback and saw the need to make the core algorithms faster, lighter, and more portable.

That’s why we teamed up with HuggingFace: to create outlines-core, a Rust implementation of the core algorithms. Together, we made it easier to integrate structure enforcement wherever inference happens.

Porting this to Rust brings:

• Speed: ~2× faster index compilation, no JIT latency on first run.
• Lightweight integration: the core is dependency-light, making it easier for other libraries (e.g. in transformers, TGI) to adopt structured generation without pulling in the full Outlines stack.

Everything you need to know about the release and how to get started, head over to the co-authored blog post by the .txt and HuggingFace teams.