NEWS

Tencent launches Hy4, model with vision and computer control, available via OpenRouter

Tencent launches Hy4, model with vision and computer control, available via OpenRouter
Image: Redação iMasters
News
Reading
3 min
Tables
1
In short

Hy4, Tencent's new 770-billion-parameter model, interprets images, documents, and controls graphical interfaces and the command line. It's already on OpenRouter, which changes the calculus for anyone evaluating AI stacks outside China.

On November 4, 2026, Tencent launched Hy4, the latest version of its flagship AI model, now capable of understanding images and operating computer interfaces directly. The announcement was reported by TechNode and places the Chinese company within a race that, until now, had been dominated by Western labs such as Anthropic and OpenAI.

According to Tencent, Hy4 can interpret charts and documents, reason over multiple images at once, and operate both command-line tools and full graphical interfaces. The model has 770 billion parameters and is already available in Tencent's own products, as well as on third-party platforms such as OpenRouter.

What Hy4 does in practice

The "computer use" capability (controlling a computer the way a human would, clicking, typing, and navigating between screens) is not exclusive to Hy4: Anthropic introduced a similar feature in Claude in 2024, and OpenAI launched Operator in 2025. The differentiator Tencent highlights is training aimed at specific workflows.

According to the company, Hy4's training prioritized three areas:

  • Office tasks (spreadsheets, documents, presentations);
  • Financial analysis;
  • Game development, including the use of Blender and CAD software.

Tencent also says it tested the model on tasks that require switching between multiple programs, which is exactly the kind of scenario where desktop AI agents tend to fail: maintaining context as focus shifts from one window to another. The company's own product teams helped train and evaluate the model against real workflows, according to TechNode.

Pricing and where to find the model

Hy4's API pricing remained the same as the preview version announced before the official launch:

ItemValue
Input tokens$0.834 per million
Output tokens$2.501 per million
Tencent Cloud TokenHub discount50% until December 31, 2026

Tencent also integrated Hy4 into two of its own products: WorkBuddy (a productivity assistant) and CodeBuddy, a direct competitor to tools like GitHub Copilot. Between November 4 and 15, the company is offering a free usage quota on both products. A cheaper model, named Hy4 flash, is planned, but the source did not confirm a launch date.

Open weights: the part that's still missing

Tencent said it will release Hy4's weights and a system card "soon," but had not published either by the time of the announcement. This matters because much of the appeal of the Hunyuan models (Tencent's AI family) has always been the availability of open versions, unlike the closed models from OpenAI and Anthropic.

Until the weights are released, anyone who wants to use Hy4 depends on the paid API or OpenRouter, in the same way as a closed model. Without the system card, there is also, for now, no public documentation on safety limits, bias, or terms of use for sensitive tasks such as interface automation, something engineering teams typically require before putting an agent to click buttons in production.

What changes for those choosing an AI stack in Brazil

For a Brazilian team evaluating AI models with agent capabilities, Hy4 arrives as one more option testable via OpenRouter, without needing to open an account with a Chinese cloud provider. That lowers the barrier to testing: it's possible to compare Hy4's cost and quality side by side with Claude or GPT in the same evaluation pipeline, since OpenRouter exposes models through a unified API.

The dollar price, however, remains subject to exchange rates and to OpenRouter's own usage fees, so the actual savings depend on how the team measures cost per task, not just per token. Another open question is the lack of information about data handling and compliance with the LGPD (Brazil's data protection law) when sending screenshots, documents, or financial data for processing by a model operated on Chinese infrastructure, something Tencent did not detail in the announcement.

The promise of open weights, once fulfilled, is the factor that matters most for anyone considering self-hosting: only with the weights published could a Brazilian team run Hy4 on its own infrastructure, with full control over where the data travels. Until then, using Hy4 for desktop automation or document analysis means depending on Tencent's API or an intermediary like OpenRouter, with the same trust limitations that already exist for any closed model.

Translated from the Brazilian Portuguese original · Read the original

Read also