Vicuna Model Explained: How It Works and When to Use It

Vicuna is LMSYS's LLaMA chat model fine-tuned on ShareGPT conversations. This covers its versions, licences, measured limits and whether it fits today.

Vicuna Model Explained: How It Works and When to Use It
Written by TechnoLynx Published on 11 Jul 2026

Vicuna is a chat assistant from LMSYS, built by fine-tuning Meta’s LLaMA on conversations users shared from ChatGPT through ShareGPT. It is an auto-regressive transformer language model. Its makers released it for research on large language models and chatbots, aimed at researchers and hobbyists. The short verdict for anyone choosing an open LLM today: Vicuna matters a great deal historically, but it is a poor default for a new production application. One 2026 comparison names Qwen 2.5 7B and Mistral 7B Instruct as the better choices for local deployment. Vicuna’s limits reflect the constraints of how it was trained.

Who built Vicuna, and how was it trained from LLaMA?

LMSYS (Large Model Systems Organization) is a research collaboration across several universities, including UC Berkeley, Carnegie Mellon and Stanford. Its first release, the Vicuna-13B model, came out in March 2023. LMSYS put the training cost of that model at around $300. That figure covers only the marginal fine-tuning step, though. Meta had already paid for the expensive part when it pre-trained LLaMA on one to 1.4 trillion tokens, depending on model size.

The recipe itself is supervised instruction fine-tuning on real conversations. LMSYS collected the shared ChatGPT conversations from ShareGPT.com, converted the HTML back to markdown, filtered out inappropriate or low-quality samples, and split long conversations into segments that fit the model’s maximum context length. LMSYS describes two training choices of its own. The loss was adjusted for multi-turn conversations and computed only on the chatbot’s own output, and the maximum context length went up from 512 tokens (Alpaca’s setting) to 2,048. The training code started from Stanford Alpaca’s, with added support for multi-turn conversations and similar hyperparameters. In the later MT-bench paper, the authors list those hyperparameters for their Vicuna-7B and 13B runs: global batch size 128, learning rate 2e-5, three epochs, sequence length 2,048.

The amount of data grew across releases. Vicuna v0 used around 70K conversations, while v1.3 and v1.5 used around 125K. So the “70K” and “125K” figures quoted for the model are both correct, for different releases.

ShareGPT data gave the model one real advantage: it learned from genuine multi-turn exchanges rather than single synthetic instructions. The same paper that described the data also found that fine-tuning on ShareGPT kept improving MMLU scores as the data grew, while a small, high-quality conversation set mostly taught the model a style GPT-4 prefers, without much gain on MMLU. That split between learning a style and learning a capability is the thread running through the rest of this article.

How does Vicuna compare with LLaMA and Alpaca?

These three are not peers. LLaMA is a family of 7B to 65B parameter models trained only on publicly available datasets. Meta presented it as a versatile base, as opposed to a fine-tuned model built for a specific task. Alpaca and Vicuna are both fine-tunes of LLaMA, trained on different data.

Stanford’s Alpaca 7B was fine-tuned on 52K instruction-following demonstrations, generated self-instruct style with OpenAI’s text-davinci-003. Vicuna replaced those synthetic instructions with real shared ChatGPT conversations and a longer context. In a preliminary evaluation with GPT-4 as the judge, LMSYS reported Vicuna-13B beat LLaMA and Alpaca in more than 90% of cases. A secondary account describes Vicuna’s answers as more natural in tone than Alpaca’s and better suited to real conversation. In practice, Alpaca showed that cheap instruction tuning worked, and Vicuna showed that the choice of data mattered a lot.

It also helps to place Vicuna in the wider field, because the term “generative AI” gets conflated with LLMs. Vicuna generates text auto-regressively, predicting one token after another. Generative modelling also covers GANs, VAEs, flow-based models and diffusion models, and those families produce images, audio and 3D assets. If a project calls for synthesised images, our comparison of GAN vs diffusion model architectures is the right starting point, and the broader overview of how generative AI models work covers the full taxonomy.

Vicuna-7B, Vicuna-13B and v1.5: telling the Vicuna model versions apart

The model varies by parameter size (7B and 13B, and in v1.3 up to 33B) and by weights version. The official FastChat notes on weight versions are the primary reference, and the table below summarises them.

Version Base model Training data How weights ship Context What changed
v0 (30 March 2023) LLaMA 1 ~70K ShareGPT conversations Delta weights applied to the original LLaMA weights 2,048 tokens First release
v1.1 (12 April 2023) LLaMA 1 Not stated in the version notes Delta weights Not stated Separator changed from ### to the EOS token </s>
v1.3 LLaMA 1 ~125K, twice the earlier amount Merged weights, no delta step Not stated Larger data set; 7B to 33B sizes
v1.5 Llama 2 ~125K Merged weights 4K, plus 16K variants via linear RoPE scaling New base model with a more permissive licence

The delta-weight format meant a v0 user had to get the original LLaMA weights separately and apply the delta before the model would run. LMSYS chose that format to comply with the LLaMA licence. The v1.1 separator change made it easier to determine when generation should stop and improved compatibility with other libraries. One consequence follows: pin the prompt template to the weights version you deploy.

Llama 2 itself was released on 18 July 2023, which sets the earliest date v1.5 could have appeared.

Can Vicuna be used commercially?

It depends on the version, and even the most permissive version leaves the data question open. Our reading follows, with the reasons, but treat it as an engineer’s reading and not a legal opinion.

Versions v0 to v1.3 sit on LLaMA 1, which Meta released under a noncommercial licence focused on research, with access granted case by case. The original Vicuna demo was described as a research preview for non-commercial use only, subject to the LLaMA model licence, OpenAI’s terms of use for generated data and ShareGPT’s privacy practices. Alpaca’s authors gave the same reasoning in their own words: because Alpaca is based on LLaMA, it inherits the non-commercial licence.

Vicuna v1.5 moves to Llama 2, and FastChat says it should be used under Llama’s model licence. The Llama 2 Community License grants a royalty-free licence to use, modify and create derivative works. Two of its conditions matter here. If your products had more than 700 million monthly active users in the calendar month before the Llama 2 release date (18 July 2023), you must request a separate licence from Meta. You also may not use Llama materials or their outputs to improve any other large language model, apart from Llama 2 and its derivatives. The second clause catches teams that plan to use a Vicuna v1.5 deployment to generate synthetic training data for a different model.

The unresolved part is the data. ShareGPT conversations are ChatGPT outputs. One secondary source argues that OpenAI’s terms prohibit using those outputs to train competing models, which creates legal ambiguity even for v1.5. Stanford gave the matching reason for Alpaca’s own restriction, saying the text-davinci-003 terms prohibit developing models that compete with OpenAI. The v0, v1.3 and v1.5 model cards describe research as the primary use. We would not ship v1.5 in a commercial product until your counsel has looked at the data-provenance question as well as the base licence.

Where Vicuna’s weaknesses come from

LMSYS was forthcoming about the model’s limitations. It said the model is not good at reasoning or mathematics, may not identify itself correctly or ensure its outputs are factually accurate, and has not been sufficiently optimised for safety, toxicity or bias. On MT-bench, Vicuna-13B lagged significantly in extraction, coding and maths, and the MT-bench paper describes it as noticeably worse than GPT-3.5 and GPT-4 in reasoning, maths and coding. The same paper also found weak instruction following: used zero-shot as a judge, Vicuna-13B had consistency rates of 11.2% to 16.2% across prompt templates and error rates of 22.5% to 78.8%.

One account ties these weaknesses to the training data’s limited coverage. Gudibande et al. (2023) fine-tuned a series of models to imitate ChatGPT and found that imitation closes little to none of the gap to ChatGPT on tasks the imitation data does not heavily support. The paper also explains why this slips past human raters: imitation models copy ChatGPT’s style well but not its factual accuracy. Its conclusion is that bridging the gap takes either an unwieldy amount of imitation data or a more capable base model. One secondary account says Vicuna’s ShareGPT data gave limited systematic coverage of maths, logic and coding.

Why the “90% of ChatGPT” figure needs care

LMSYS had GPT-4 score answers to 80 questions and summed the scores, which gave Vicuna a total of 92% of ChatGPT’s. In the same evaluation, GPT-4 rated Vicuna’s answer better than or equal to ChatGPT’s on only 45% of the questions. So the 92% is an aggregate of relative scores. It does not mean the model matched ChatGPT nine times out of ten. LMSYS itself described the evaluation as preliminary and noted possible position, verbosity and self-enhancement biases in a GPT-4 judge. A later secondary review adds that the question set was not a standardised benchmark. How to read LLM-as-judge results properly is a measurement question in its own right. Here, the practical point is that the headline score reflects style more than capability.

GPU memory needed to self-host the 7B and 13B models

According to the FastChat repository, its standard serving command needs about 14GB of GPU memory for the 7B model and about 28GB for the 13B model. Those numbers match weights held in 16-bit precision at two bytes per parameter (7 billion × 2 bytes ≈ 14GB), before any extra room for the KV cache on long contexts. Weight quantisation lowers the VRAM bar and makes LLaMA-family models usable on less powerful hardware. A 2023 cost-performance study of five open models chose Vicuna-7B for an email-automation setting on device-constrained hardware and Vicuna-13B for game NPC dialogue under VRAM constraints. Treat those recommendations as reflecting the models available at the time. For the self-hosting details (serving stacks, quantised builds, GPU sizing), our guide on when to self-host Vicuna on GPU covers what this section leaves out.

Does the Vicuna LLM still make sense for a new application?

For new production work, our answer is generally no. One 2026 comparison reports roughly 50% MMLU for Vicuna 7B v1.5, against about 60% for Mistral 7B Instruct and 74% for Qwen 2.5 7B, and says Llama 2 7B Chat performs similarly to Vicuna on most metrics. Those figures are consistent with the structural argument, though: a 2023 imitation fine-tune of an older base will not match newer bases trained with more and better data. Meta’s own Llama 2 Chat added reinforcement learning from human feedback, rejection sampling and PPO on top of supervised fine-tuning, while Vicuna v1.5 used supervised instruction fine-tuning.

Fine-tuning Vicuna v1.5 still leaves you under Llama’s licence and its data-provenance question. Putting the same effort into a current base model usually gets you further.

A quick check before choosing Vicuna

  • Are you reproducing 2023 research or studying ShareGPT-style training? Vicuna is the right artefact, and this is the use its model cards describe.
  • Are you maintaining a system already built on it? Pin the weights version and prompt template, and plan the migration rather than reaching for a newer release.
  • Does the task depend on maths, coding or extraction? These are the measured weak spots, so look elsewhere.
  • Does the task depend on factual accuracy? The base model will not supply it. Grounding answers in your own documents through retrieval augmented generation matters more than which chat fine-tune you pick.
  • Is the product commercial? Only v1.5 sits on a base licence that permits commercial use, with conditions, and the ShareGPT data question still applies.
  • Do you expect to swap models later? Design the integration so the model is replaceable. Our piece on LLM-agnostic architecture for model procurement explains why that matters more than any single model choice.

Vicuna’s lasting contribution is less the model than what it showed: real conversational data shapes a chatbot more than synthetic instructions do, and style is cheap to learn while capability is not. For teams weighing open models against a specific product need, TechnoLynx’s generative and agentic AI R&D work evaluates model provenance, licence chain and task fit. If a newer chat fine-tune claims to rival a proprietary model, ask which base model it starts from.

Back See Blogs
arrow icon