FastChat

FastChat is an open platform for training, serving and evaluating large language models, with 39,000+ stars on GitHub — best known as the project behind Vicuna and Chatbot Arena.

Overview

Created by LMSYS, the research group spun out of UC Berkeley and partner universities, FastChat is the infrastructure that powered Chatbot Arena (now LMArena), the crowdsourced benchmark where models battle side by side: the platform has served over 10 million chat requests across 70+ LLMs and collected more than 1.5 million human votes for its Elo leaderboard. Inside the repository you get the full loop in one place — training code for models such as Vicuna, a distributed multi-model serving system with a web UI and OpenAI-compatible REST APIs, and evaluation tooling including MT-Bench for multi-turn chat quality. That combination made FastChat a standard reference stack in the open-model ecosystem: if you want to serve several open models behind one interface and compare them honestly, this is one of the original toolkits for it. The repository remains maintained, with updates through 2026.

Key Features

• Multi-model serving – distributed controller/worker system serving many models behind one web UI.
• OpenAI-compatible APIs – REST endpoints that existing chat clients and code can talk to directly.
• Chatbot Arena heritage – the side-by-side battle and Elo-rating methodology from LMArena.
• Vicuna training code – the recipes behind the Vicuna family of open chat models.
• MT-Bench evaluation – multi-turn question sets for scoring chatbot quality.
• Many model backends – serve Hugging Face models and accelerated engines from one platform.
• Battle-tested at scale – proven serving millions of public chat requests.

💳 Pricing

FastChat is completely free and open source under the Apache License 2.0. You host it yourself, so costs are limited to your own servers or GPU instances.

Who Is It For

Researchers and engineering teams who serve multiple open-weight models, run their own evaluations or leaderboards, or need a reference implementation of arena-style human preference testing.