OpenChat 3.5

OpenChat 3.5 is a 7-billion-parameter open model fine-tuned from Mistral 7B using C-RLFT, a strategy inspired by offline reinforcement learning that learns from mixed-quality data without preference labels. Released in November 2023, it was the first 7B model to score comparably with ChatGPT on MT-bench, hitting 7.81 and topping the open-source leaderboard at the time.

The project emphasizes commercially viable open models: code and weights are distributed under Apache 2.0. A December 2023 update further improved coding performance. OpenChat 3.5 became a widely used base for downstream fine-tunes and local chat deployments.