Vicuna

Vicuna was one of the defining open chat models of the early post-ChatGPT era. Built by LMSYS, the team behind Chatbot Arena, it fine-tunes Meta's Llama and later Llama 2 on roughly 125,000 user-shared ChatGPT conversations from ShareGPT, and early GPT-4-judged evaluations placed it near 90 percent of ChatGPT quality, a result that helped legitimize instruction-tuned open models. Weights come in 7B, 13B, and 33B parameter sizes with context variants up to 16K tokens. The model lives inside FastChat, LMSYS's open platform that provides its training code, a distributed multi-model serving system with a web UI and OpenAI-compatible REST APIs, and the MT-Bench evaluation harness that scores chatbots using LLM judges. FastChat itself is Apache 2.0, while Vicuna weights follow the underlying Llama license, so commercial terms track Meta's. Though newer models have surpassed it, researchers still cite Vicuna as a baseline and use FastChat as serving and evaluation infrastructure.