Tencent is pricing Hy4 for enterprise agents rather than benchmarks, so adoption inside WorkBuddy and on OpenRouter will decide whether its heavy AI spending pays back.
Tencent released the official version of Hy4, its flagship large language model, on November 4, adding the ability to read images and to operate a computer, IT Home reported. The company says the model puts its Hy series among the strongest open-source models for the first time. It also concedes that Hy4 still trails the leading closed models by a wide margin, especially on Tencent’s own hard internal tests of frontier math, long development tasks, model iteration and visual question answering.
The release follows the Hy4 preview of August 28, which had 770 billion total parameters, about 49 billion of them active for each request, and a context window of 1 million tokens. Tencent says it has significantly scaled up both pre-training and post-training since then, and that performance is markedly better than the preview while the API price stays the same. The model is live on Tencent Cloud and OpenRouter, and users can try it in the CodeBuddy and WorkBuddy agent tools, the Yuanbao assistant and the ima knowledge app.
Two parts of the package remain to come. Tencent says it will open-source the model weights soon and publish a system card with fuller safety evaluations, but it gave no date for either. A smaller Hy4 flash version, aimed at lower cost and faster responses, is also planned.
Computer use and vision move Hy4 toward office work
Hy4 is the first model in Tencent’s Hy series that can operate a computer, mixing command-line instructions with clicks in a graphical interface. Tencent ran dedicated training for Blender, the 3D modeling program, for computer-aided design (CAD) software, and for long tasks that move between applications, according to IT Home. The vision features combine perception, reasoning and tool calls, so the model can process fine detail in images, analyze complex charts and solve visual math problems.
Tencent lists the target uses as visual and computer tasks, front-end and game development, finance and office work, research and model development, and long-running personal assistance. Caixin Global reported that Hy4 was developed alongside WorkBuddy, Tencent’s agent product for office tasks, with an emphasis on practical jobs and reliability, and described the launch as part of Tencent’s push to embed advanced AI in enterprise software.
That framing explains the pricing decision. Keeping the API price unchanged while raising capability gives developers and companies that built on the preview a direct upgrade path, and lets Tencent’s own agent products move to the stronger model without raising their running costs.
Tencent is opening the weights more cautiously than before
The preview was released as an open model in August. For the official version, Tencent says it wants a more responsible approach to open-sourcing, with particular attention to cybersecurity and biological risks, and it has not yet published the weights. It says the system card it plans to release will carry more complete details of its safety evaluations.
To make deployment easier once the weights are out, Tencent has worked with the vLLM and SGLang open-source serving communities and with Inferact Inc., so that the model can run on common open-source infrastructure.
Tencent says it will keep scaling pre-training and post-training to close it, but it published no third-party benchmark to show where Hy4 now stands against DeepSeek, Moonshot AI’s Kimi or Alibaba’s Qwen. EastFrontier reported in August that Tencent released the Hy4 preview for China’s AI agent race, and the official release keeps that agent focus while adding the tools agents need to work inside desktop software.
The weights and the system card are the next markers
The most concrete thing to watch is the open-weight release. When the weights and system card appear, outside researchers will be able to test Hy4’s coding, vision and computer-use claims and check how the model handles the cyber and bio risks Tencent says it screened for.
The second marker is Hy4 flash. A cheaper, faster version would matter more for Tencent’s agent products than the flagship, because tools such as WorkBuddy run many model calls for each task. Usage rankings on OpenRouter in the weeks after release will give an early, public read on whether developers are choosing Hy4 over rival Chinese models at the same price.