Guanaco
The 2023 research chatbots that demonstrated QLoRA 4-bit fine-tuning
Overview
Guanaco is a family of 7B to 65B chatbots fine-tuned from LLaMA on the OASST1 dataset, released in May 2023 by University of Washington researchers to demonstrate QLoRA 4-bit fine-tuning. The weights remain on Hugging Face for research use under the LLaMA licence terms, but the project has been dormant since July 2023.
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 These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete.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 Guanaco
- The QLoRA paper and MIT-licensed training code for 4-bit fine-tuning with LoRA adapters
- Guanaco adapter weights for LLaMA 7B, 13B, 33B and 65B on Hugging Face
- Merged full-weight Guanaco 13B, 33B and 65B checkpoints
- Scripts reproducing the Guanaco training run on the OASST1 dataset
- A worked example of fine-tuning a 65B model on a single 48GB GPU
- A documented caution on GPT-4-judged chatbot benchmarks
Guanaco is worth knowing for the method, not the model. QLoRA showed that 4-bit quantisation plus LoRA adapters could fine-tune a 65B model on one 48GB GPU, and that machinery now lives in bitsandbytes and Hugging Face PEFT. The headline 99.3%-of-ChatGPT score was a GPT-4-judged result on a small benchmark the paper itself warned was unreliable.
The chatbots are a 2023 artefact: LLaMA-based, research-only, unmaintained since July 2023 and without any hosted service. Read the paper and use QLoRA through a current fine-tuning stack; download Guanaco only to reproduce the research.
What we found
Guanaco is a set of chatbot models released in May 2023 alongside QLoRA, a paper by Tim Dettmers, Artidoro Pagnoni, Ari Holtzman and Luke Zettlemoyer at the University of Washington. The paper later appeared at NeurIPS 2023. Its method is the reason to know the name: freeze a base model in 4-bit precision using a new NormalFloat data type, add double quantisation and paged optimisers to save memory, and train small LoRA adapters on top. According to the paper, that brought fine-tuning a 65-billion-parameter model down to a single 48GB GPU without losing 16-bit task performance.
Guanaco was the demonstration. The authors tuned LLaMA 7B, 13B, 33B and 65B on the OpenAssistant OASST1 dataset and reported that the best model reached 99.3% of ChatGPT's level on the Vicuna benchmark after 24 hours on one GPU. That figure came from GPT-4 acting as judge on a small prompt set against the March 2023 ChatGPT, and the same paper cautioned that current chatbot benchmarks were not trustworthy, so it is best read as a measure of the method rather than the chatbot.
This is a legacy page. The qlora repository is MIT-licensed and still online but dormant, with its last commit in July 2023 and nearly 200 open issues. The adapter weights, plus merged 13B, 33B and 65B checkpoints, remain downloadable from Hugging Face. The model card puts the adapters under Apache 2.0 but says Guanaco must be used according to the LLaMA licence, which restricted the original LLaMA to non-commercial research, and the README calls the models purely for research. There is no hosted service, API or support.
The method outlived the models. QLoRA's 4-bit machinery lives on in bitsandbytes, which remains MIT-licensed and actively maintained, with version 0.50.2 released in August 2026, and in Hugging Face PEFT. Anyone wanting cheap fine-tuning today should use those libraries with a current base model rather than Guanaco itself.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- Guanaco 65B was fine-tuned in 24 hours on a single GPU that costs $2/hour to rent — demonstrating that frontier-class model fine-tuning is accessible to independent researchers.
- The QLoRA technique developed for Guanaco reduces the memory required to fine-tune Llama 2 70B by 75% compared to full precision training.
- The QLoRA paper is one of the most cited machine learning papers of 2023 and directly enabled consumer-friendly fine-tuning tools like Axolotl and LLaMA Factory.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Guanaco's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Guanaco 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 Guanaco directly →
Spotted something out of date? Suggest an update →
Related reading
Science
Eighteen Years of Whale Calls Show Rare Omura's Whales Favour Warm, Food-Rich Waters off A...
Cybersecurity
CISA Orders Three-Day Fix for Exploited Collabora Flaw in Nextcloud's Built-In Office Serv...
AI & ML
DeepSeek Open-Sources DeepInfer, the Engine That Serves Its V4.1 Flash Model on NVIDIA and...
Alternatives to Guanaco
More in LLMs & Chat
