Skylark (Yunque)

What it is

Skylark — Yunque (云雀) in Chinese, the word for skylark — is ByteDance's first large language model. It debuted in August 2023 under the name Skylark and was later marketed as Doubao.^[1] Caixin reported that the model was originally called Yunque and that it received algorithm registration in 2023.^[2] It was among the first Chinese large language models to pass the national algorithm registry.^[1]

A large language model is a neural network trained on very large amounts of text to predict what comes next; scaled up, that training yields a system that can answer questions, write and summarise. The registry is the filing that providers of generative AI services complete with Chinese regulators before a model is offered to the public, so appearing on the list early is part of what makes a public launch possible.

How it came about

Skylark debuted in mid-August 2023, later than Baidu's Ernie Bot, which appeared in March 2023, and later than the models from Huawei and Alibaba that followed Ernie Bot.^[4] ByteDance was therefore not the first Chinese company to put a large model in front of users; its first entry came several months after the first wave of competitors.

What it powered

The model powered the Doubao chatbot, which opened for invitation testing in August 2023.^[3] By March 2024, Jiemian reported that Skylark, first released in August 2023, had become the base beneath Doubao and other ByteDance AI products.^[4] The same report said a major version upgrade of Skylark was due in April 2024,^[4] and that the model's still-immature performance had partly slowed the rollout of ByteDance's AI applications.^[4]

Scale of use

By May 2024 the model was handling an average of 120 billion text tokens and generating 30 million images a day, and had been integrated into more than 50 ByteDance businesses, including Douyin, Fanqie Novel, Feishu and Oceanengine.^[1] These are usage figures rather than quality figures: tokens measure how much text the model processes and the image count measures how much it draws, and together they show how widely one model had been wired into a company's existing products.

Relaunch as Doubao

On 15 May 2024 Volcano Engine, ByteDance's cloud unit, relaunched the model publicly as the Doubao model family.^[1] The rename folded the model and the assistant it runs under a single brand: Doubao became the name both of the family of models offered to developers and, since the invitation testing of August 2023, of the chatbot that ordinary users could try.^[1][3]

Where it sits

Skylark is the starting point of ByteDance's model line. It appeared in August 2023 as a single model, first sat behind an invitation-only chatbot, then spread across more than 50 internal businesses, and was finally renamed and offered publicly as a model family through the company's cloud platform.^[1][3] The public record of its first year also shows the limits of that path: an early version upgrade was planned and immature performance held back some application rollouts, so the model's history is one of internal deployment first and external packaging second.^[4]