OpenChat 3.6

OpenChat 3.6 8B (openchat-3.6-8b-20240522) is an instruction-tuned language model from the OpenChat team, released in June 2024 and built on the Llama 3 8B architecture. It is fine-tuned with C-RLFT (Conditioned Reinforcement Learning Fine-Tuning), a strategy inspired by offline reinforcement learning that trains on mixed-quality data without preference labels.

The Apache 2.0 weights support an 8,192-token context and were reported to outperform Llama 3 8B Instruct and comparable open fine-tunes on general conversation, coding assistance, and mathematical reasoning, making it a strong drop-in open chat model for its era. Date is era-level (June 2024).