Subscribe
China
4 min read

Doubao Work Adds Kimi, GLM and DeepSeek Models

ByteDance's Doubao Work now lets Team and Enterprise customers use Kimi K3.1 Pro, GLM-5.3 and DeepSeek V4.1 Flash, billed from existing AI points.

Olivetti-style cover reading DOUBAO WORK with three model blocks feeding one chat window on cream paper
Olivetti-style cover reading DOUBAO WORK with three model blocks feeding one chat window on cream paper

ByteDance's office AI product, Doubao Work, now lets business customers use models from three Chinese rivals inside the same chat window. On October 29, Doubao Work said it had opened third-party model services to enterprise customers, with Moonshot's Kimi K3.1 Pro, Zhipu's GLM-5.3 and DeepSeek V4.1 Flash in the first batch, Leiphone reported. The statement names "Kimi 3.1 Pro", "GLM 5.3" and "DeepSeek V4.1 Flash" among the first batch of model series to be added. It does not say which models come next.

For a company, the change is about procurement more than technology. There is no new contract with each lab and no separate API bill. The models draw on the same pool of AI points the company already pays ByteDance for.

What changed

According to the reports from Leiphone, Jiemian and National Business Daily, which all carry the same company statement:

  • Who gets it: customers on the Doubao Team subscription and the Doubao Enterprise subscription.
  • On by default: the feature is switched on when it launches. Companies do not need to buy from the model vendors or file a separate application.
  • Billing: use is deducted from the company's existing AI points, based on usage.
  • Who can use which model: admins can open models to specific members, departments, groups or the whole company. They can turn on all models from one vendor in one step, or configure a single model on its own.
  • Where it works: admins can choose which surfaces each model appears in: the Doubao desktop app, the Doubao Work desktop app, Doubao mobile, Doubao on the web, and Doubao Work inside Feishu, ByteDance's office suite (sold abroad as Lark). Leiphone gives the example of a company that allows a model only on "work computer" and the Feishu workbench, to balance business efficiency with data security.
Diagram: Kimi K3.1 Pro (Moonshot), GLM-5.3 (Zhipu) and DeepSeek V4.1 Flash feed the Doubao Work admin console, which sets access by member, department, group or company and by surface: Doubao desktop, Doubao Work desktop, Doubao mobile, Doubao web and Doubao Work in Feishu

What was possible before

This is not the first time ByteDance's office tools could reach outside models. Doubao Work Partner, the agent product renamed from Feishu aily on August 14, 2026, already supports custom model access. A company connects model APIs it has opened with Volcano Engine, Alibaba Cloud or Tencent Cloud, such as Doubao, DeepSeek or Kimi, and pays the model vendor directly, while Doubao Work Partner's output and tool fees still use AI points. Since June 1, 2026, this option is limited to higher paid tiers, and overseas models are not supported (Feishu help center).

The new step puts the three rival models directly in Doubao Work's model menu, with no vendor account to set up and one bill in AI points.

What it costs

ByteDance has not published how many points each third-party model costs. Its existing price lists, as of October 2026, give the frame:

  • Doubao Team subscription: 1,988 yuan per seat per year on annual billing (shown as 166 yuan a month), or 198 yuan per seat per month on monthly billing, with 2,000 AI points per seat per month, from one seat, on the Doubao Work pricing page. The help center says it includes an AI quota. The pricing page also says enterprise data is not used for model training.
  • Doubao Enterprise subscription: sold from 100 seats through sales only. Seats include no AI quota, so the company must also buy a usage package (help center).
  • Usage packages: the Feishu billing page lists 2 million points for 99,000 yuan a year, 6 million for 297,000 yuan and 20 million for 990,000 yuan (Feishu help center).

For the related Doubao Work Partner agent, that same billing page says model fees are charged by converting each vendor's list price into AI points. Whether Doubao Work uses the same rule for Kimi, GLM and DeepSeek is not stated in the October 29 announcement.

Bar chart of Doubao enterprise usage packages, October 2026: 2 million points for 99,000 yuan a year, 6 million for 297,000 yuan and 20 million for 990,000 yuan; Team subscription 1,988 yuan per seat per year or 198 yuan per seat per month

Why it matters

All three models have been in the news this month, and our guide to China's AI labs explains who makes each one. We covered Kimi K3.1 Pro, the DeepSeek V4.1 Flash 1017 update and Irregular's cyberattack tests of GLM-5.3. All three labs sell their own chat apps and APIs. Putting their models inside a ByteDance product gives them a route to Feishu's business users, while ByteDance keeps the customer, the bill and the admin console.

For ByteDance, the bet is that the workspace matters more than the model. If an employee can switch between Doubao, Kimi, GLM and DeepSeek in one window, the reason to stay is the files, chats, permissions and audit tools around the chat, not the model behind it.

The timing is notable. On the same day, Tencent launched Buddy One, which puts its office apps behind one account and one credit pool. At that event, Tencent's Tang Daosheng said of models, agents and chat systems that "the choice is in the enterprise's hands," the Economic Information Daily reported. ByteDance and Tencent now sell office AI on a similar pitch: one credit balance, a choice of models, and admin control over who uses what.

There are also open questions. The announcement does not say how data sent to a third-party model is handled, whether those prompts leave ByteDance's systems, or what each model costs in points. Companies that care about the recent report that DeepSeek and Moonshot distilled OpenAI's GPT-6 Astra, or about where their data goes, will want those answers before they switch the models on for the whole company. Because the feature is on by default, admins who do not want it should check their settings.

Sources

ChinaNewsEnterprise AILLM