OpenPipe

Why we included this project

Teams that send thousands of similar requests to large frontier models often end up paying per-token rates for repeated work. OpenPipe gives you a different path: log your real traffic through its OpenAI-compatible endpoint, then use that logged data to fine-tune smaller, cheaper models like Mistral and Llama 2, and swap between them by changing a single model name. It also includes an evaluation step that compares a fine-tuned model's output against the base model on your own logs, so you can check quality before you commit. One honest caveat for self-hosters: the maintainers paused work on this open-source release to integrate proprietary code, so treat the current snapshot as a proven reference architecture rather than an actively maintained codebase.