FastChat
The open platform for training, serving, and evaluating LLMs: home of Vicuna and the Chatbot Arena side-by-side battle UI.
About
FastChat is the LM-SYS (Berkeley) open platform for everything around large language models: training, serving, and evaluation in one repo. It is where Vicuna was built and released, and its serving stack, a controller plus model workers plus a web GUI, became the template for open LLM deployment. The headline feature is the Chatbot Arena: a side-by-side battle UI where two anonymous models answer the same prompt and users vote, the mechanism that produced the LMSYS leaderboard everyone cites.
Beyond the arena it offers OpenAI-compatible REST APIs and SDKs, Hugging Face generation API integration, LangChain integration, and fine-tuning code with data and hyperparameters for reproducing Vicuna-class models, including LoRA support and SkyPilot scripts for any-cloud training. At 39,500+ stars and Apache-2.0 it is one of the most-starred LLM repos ever. The honest note is cadence: the last push was May 2026, and much of the ecosystem's energy has moved to vLLM/SGLang for serving, so treat FastChat as the battle-tested classic rather than the bleeding edge.
What it can do
- Full serving stack: controller, model workers, web GUI, and CLI inference
- Chatbot Arena side-by-side battle UI behind the LMSYS leaderboard
- OpenAI-compatible RESTful APIs and SDKs
- Fine-tuning code, data recipes, and hyperparameters (Vicuna, LoRA)
- Evaluation tooling for LLM benchmarking
- SkyPilot integration for fine-tuning on any cloud
Setup
- Install:
pip install "fschat[model_worker,webui]"(or install from source per README). - Download model weights (Vicuna deltas or any supported HF model) following the Model Weights section.
- Serve: launch the controller, model worker(s), and web GUI, or use the CLI inference directly.
- For APIs, point clients at the OpenAI-compatible REST endpoint.
- For the arena, follow the 'Launch Chatbot Arena' section to run the side-by-side battle UI.
Prerequisites
Python 3.8+; GPU(s) for serving and training; model weights from the Hugging Face Hub (some require accepting a license).
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: 39,548 GitHub stars (GitHub API), Apache-2.0; the README documents the serving stack, the arena launch, the OpenAI-compatible APIs, and the fine-tuning recipes; LM-SYS/Berkeley provenance is well established. Strengths: historic, enormous adoption, the arena mechanism the whole field uses, complete train/serve/eval loop. Docked for: (1) pushed May 2026, slower cadence than peers; (2) serving mindshare has shifted toward vLLM/SGLang; (3) sprawling repo, setup has many moving parts. License: Apache-2.0.
★ 39,548 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 9: starScore from 39,548 stars on lm-sys/FastChat (log scale vs 298k peer max)
- · Maintenance 6: pushed 2026-05-01 (recency ~4); slower cadence than peers
- · Openness 10: Apache-2.0 (0.7*10) + extensive README docs verified (0.3*10)
- · Manual 8: historic train/serve/eval platform, Chatbot Arena mechanism corroborated by README; docked: serving mindshare shifted to vLLM/SGLang
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.