Hugging Face launches HuggingChat

Background

A chatbot is a program that holds a conversation in ordinary language. Once large language models became good at following instructions, chatbots built on top of them could do more than answer questions: they write and explain code, draft messages, and produce text on request. Most of the best-known ones have been closed products, reachable through a company's website but with the model underneath kept private.

Hugging Face, a company that hosts open models and datasets and builds tools around them, moved into that space with a product of its own. Its co-founder, Julien Chaumond, described the launch in exactly those terms.

What happened

On 25 April 2023 Hugging Face launched HuggingChat, a free chatbot at hf.co/chat ^[1]. The company pitched it as an open-source alternative to ChatGPT ^[1]. Chaumond called it the first open-source alternative to ChatGPT ^[3].

The model behind the chat window came from another project. HuggingChat ran OpenAssistant's latest LLaMA-based model ^[2]. Open Assistant was a volunteer project organised by LAION ^[3]. In other words, the interface was Hugging Face's; the model it served had been trained by a distributed group of volunteers rather than by a single company, and it was built on top of LLaMA, an existing family of language models.

Key details

HuggingChat could be used in two ways. It was available through a web interface, and also through Hugging Face's API, so that other apps could call it and build it into their own products ^[4]. That second route matters for how open tools spread: a chat website is something a person visits, while an API is something a developer can put inside a different piece of software.

On capability, TechCrunch reported that it could handle many of the tasks ChatGPT handled, such as writing code, drafting emails and composing rap lyrics ^[4].

Open Assistant's own stated goal was an assistant small and efficient enough to run on consumer hardware ^[4]. That is a different ambition from raw scale. A model that fits on ordinary hardware can be run by whoever has that hardware, without depending on a company's servers, which is part of what the open-source framing was about.

Why it mattered

The launch put a free, openly presented chatbot behind a simple public address at a company already known for hosting open models. It gave the open side of the field a visible front door, and it gave a volunteer project a large audience: the model that Open Assistant's contributors had built was suddenly the engine of a product anyone could try.

It also made the contrast between the two approaches easy to see. One chatbot was a closed product from a single company; the other was an interface wrapped around a model trained by volunteers, offered for free and reachable both by people and by other programs. HuggingChat did not settle that contrast, but it made it concrete, and it put Hugging Face's name on the open side of it.