Vicuna
LMSYS's 2023 LLaMA fine-tune that led to FastChat and Chatbot Arena
Overview
Vicuna is an open chat model from LMSYS, first released in March 2023 by fine-tuning Meta's LLaMA on about 70,000 ShareGPT conversations. Versions through v1.5 (August 2023, on Llama 2) remain downloadable from Hugging Face in 7B, 13B and 33B sizes. It is no longer developed and is kept as a dated research model.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with Vicuna
- Downloadable 7B and 13B chat weights, plus 16K-context v1.5 variants and a 33B v1.3
- A documented low-cost fine-tune: about $300 of training for the 13B model on 70K ShareGPT conversations
- FastChat, the Apache-2.0 code for training, serving and evaluating the models
- A case study in GPT-4-as-judge evaluation, which its authors labelled non-scientific
- The origin of Chatbot Arena, now run by Arena at arena.ai
- A research baseline only, with no hosted API and no further versions
Vicuna showed in March 2023 that a LLaMA fine-tune on shared ChatGPT conversations could make a capable chatbot for about $300 of training, and the project gave the field FastChat and Chatbot Arena, now run by the company Arena. The weights through v1.5 remain on Hugging Face.
As a model it is dated: no hosted API, no release since 2023, non-commercial terms before v1.5, and training data drawn from ChatGPT outputs. Study it for the history; deploy something current.
What we found
Vicuna is a chat model released on 30 March 2023 by LMSYS, a research group spanning UC Berkeley, CMU, Stanford, UC San Diego and MBZUAI. The team fine-tuned Meta's LLaMA on about 70,000 conversations that users had shared from ChatGPT on ShareGPT.com, trained on eight A100 GPUs in a day, and put the cost of training the 13B model at about $300. The headline, "more than 90%* quality of OpenAI ChatGPT", carried its own asterisk: the blog called it "a fun and non-scientific evaluation" in which GPT-4 judged answers to 80 questions.
The versions that followed are what can still be downloaded. v1.1 in April 2023 changed tokenisation; v1.3 in June doubled the ShareGPT data and added a 33B model; v1.5 in August 2023 moved to Llama 2 and added 16K-context variants. The early versions inherit LLaMA's non-commercial terms, while v1.5 carries the Llama 2 Community License. All sit ungated on Hugging Face; the 13B v1.5 repository was last updated in March 2024, and no later version has appeared. FastChat's README says the 7B model needs about 14GB of GPU memory and the 13B about 28GB.
Vicuna's longer legacy is the tooling around it. FastChat, the Apache-2.0 training and serving code released with it, still receives commits, though its last tagged release was in February 2024. Chatbot Arena, the pairwise human-vote leaderboard LMSYS built to compare chat models, is listed by LMSYS as graduated and now runs as a separate company, Arena, at arena.ai. LMSYS's recent work centres on systems projects such as the SGLang serving engine.
As a model Vicuna is dated: a 2,048-token context in its original form, training data drawn from shared ChatGPT conversations, and no hosted API. It stays listed because the weights remain downloadable. Use it to study how open chat models began, not to deploy.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- Vicuna was trained in one day for approximately $300 — 100,000x cheaper than the estimated cost of training the original ChatGPT.
- The name 'Vicuna' continues the South American camelid theme started by Meta's Llama model.
- Vicuna's release sparked the LMSYS Chatbot Arena, a crowd-sourced human evaluation benchmark where users compare AI models anonymously.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Vicuna's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Vicuna unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Vicuna directly →
Spotted something out of date? Suggest an update →
Related reading
Alternatives to Vicuna
More in LLMs & Chat
