01.AI launches Yi-Large and the Wanzhi app

Background

By the spring of 2024 01.AI was holding its first-anniversary event, and the launches it presented were the company's most detailed public statement about its models to that date ^[1]. A first anniversary is a natural moment for a company of this kind to present its next models to the public.

Chinese companies building large language models were at this point competing on two fronts: releasing models that developers could download and use freely, and matching the best closed models from abroad on public tests. 01.AI's event addressed both fronts.

What was launched

At the event on 13 May 2024, 01.AI launched Yi-Large, a closed model with about one hundred billion parameters ^[1]. Parameters are the adjustable numbers that a model learns in training, so the figure is one common way of describing a model's size.

The company also open-sourced the Yi-1.5 series in 34B, 9B and 6B sizes. Yi-1.5 was described as an upgrade of the original Yi models, with base and chat versions in each size ^[3].

It also launched Wanzhi, a free bilingual AI workspace available on the web and as a WeChat mini program. A mini program is a lightweight application that runs inside the WeChat app. Wanzhi was described as a productivity tool for both Chinese and English users ^[1][3].

Finally, 01.AI opened its API platform worldwide. An API platform gives developers a standard way to send requests to a model, so the same service can be used from software built anywhere in the world ^[1][3].

Key details

The company reported that Yi-Large ranked second on the length-controlled win rate of AlpacaEval 2.0, behind GPT-4 Turbo, and that it ranked first among Chinese models on SuperCLUE ^[1][2]. These rankings were the main evidence the company offered for the model's quality.

The Yi-Large API was priced at 20 yuan per million tokens. The company described this as about one third of GPT-4 Turbo's price ^[2]. Price matters in practice because it decides how much an application built on the model costs to run.

The company also announced that it had started training Yi-XLarge, a Mixture-of-Experts model aimed at GPT-5-level performance ^[3]. The announcement was a statement of intent rather than a release, so readers should not treat it as a model that was available at the time.

Finally, 01.AI said that its overseas productivity apps had nearly 10 million users in total ^[1]. This is a company-reported figure and was given as a combined total across its overseas apps.

Impact

The event showed the company's approach to the market. It offered an open line of models for developers and researchers, a closed flagship for paying customers, and a consumer app for everyday tasks. Each offering reached a different group of people, and together they gave the company a presence in several markets at once.

For readers who are new to AI, the event also illustrates how a model's quality is claimed and checked. Most of the claims came from the company itself, but the leaderboards named in them, such as AlpacaEval and SuperCLUE, are public, so people outside the company could test the claims. Reading a launch with that in mind helps separate what a company says about a model from what independent tests show.

The launch was followed by later releases from the same company, but this entry covers only the May 2024 event.