OpenChat 7B

OpenChat 7B is an open-source language model fine-tuned with C-RLFT, a technique inspired by offline reinforcement learning that learns from mixed-quality data without preference labels. Built on the Mistral 7B foundation, it delivers results comparable to ChatGPT despite its smaller 7B parameter count, scoring MT-Bench 7.81, MMLU 64.3, and GSM8k 77.3. The model handles general conversation, multi-turn dialogue, and reasoning tasks, with a dedicated coding mode for software development assistance. Its Apache 2.0 license makes it commercially deployable, and it is light enough to run on a single consumer GPU such as an RTX 3090. OpenChat 7B is engineered for production use with an OpenAI-compatible API server, and its training uses padding-free techniques and the Multipack Sampler that achieve a 3 to 10x speedup over conventional padded methods. It demonstrates that high-quality chat performance can be reached at the 7B scale without the overhead of much larger alternatives.