Collaboration releases Vicuna – environmentally friendly, cost-effective rival to ChatGPT

Large Language Models (LLM) have risen to global prominence thanks to a chatbot developed by OpenAI known as ChatGPT, which can generate rich, lifelike written content and imitate human language style. Unfortunately, building and maintaining such chatbots is both financially and environmentally unsustainable, with compute costs running into the tens of millions of dollars per year, and a carbon footprint akin to that of a small city.

A global collaboration between researchers at MBZUAI, UC Berkeley, CMU, Stanford, and UC San Diego set out to address the unsustainable costs of chatbots with the creation of Vicuna — an open-source chatbot that cost $300 USD to train and has a tiny carbon footprint compared to ChatGPT, which cost over four million dollars to train with an estimated 500 tons of carbon emitted. Vicuna is also highly competitive with ChatGPT and Google’s Bard scoring around 90% on subjective language assessments and making it the strongest known, publicly available chatbot alternative.

With a minuscule size-to-power ratio, Vicuna can be easily accommodated into a single GPU accelerator, compared to the dozens of GPUs required by ChatGPT. Not surprisingly, Vicuna’s accessibility has made it an overnight sensation. On the open-source website, Github Vicuna has earned over 12,000 stars in only two weeks. Stable Diffusion, by contrast, took 15 months to reach 50,000 stars.

Vicuna was named after a relative of the llama. It was created using Meta AI’s LLaMA model (released in February 2023) as a base and then fine-tuned on user-shared conversations. Vicuna delivers responses reaching 90% of the subjective quality of ChatGPT, at a fraction of the energy budget and memory footprint. Best of all, it’s open source, so that everyone can use the model to help further green modes of content generation. Vicuna is publicly available at https://vicuna.lmsys.org/

Vicuna has approximately 13 billion parameters, which allows it to fit on a single modern GPU accelerator. For comparison, ChatGPT has been estimated to contain between several hundred billion to over one trillion parameters (OpenAI has never made the details public). This is significant because LLMs with more parameters have been expected to result in better outcomes, a narrative supported by OpenAI when describing their predecessor models GPT-2 and GPT-3. Vicuna’s research undermines that orthodoxy and demonstrates that quality does not have to be compromised in the interest of a much cheaper and more sustainable chatbot.