Tencent Hy4 Goes Full Release With Vision and Computer Use

Tencent Hy4 model goes from preview to full release

The Tencent Hy4 model is now out in full. Tencent announced the full release on November 4, 2026, adding vision and computer-use capabilities to a model that had previously only been available as a preview. The model is available through hosted APIs and inside Tencent products, while the model weights and a full system card are planned for a later release.

That split matters. The Hy4 preview, released in August 2026, was open-sourced. The full release is not yet open in the same way. Developers can use it today through Tencent Cloud and OpenRouter, and inside CodeBuddy, WorkBuddy, Yuanbao and ima, but anyone waiting to download and run the weights themselves will have to wait for the next step.

What is new in the full release

Two additions headline the release: vision and computer use. Tencent says Hy4 can now interpret fine-grained image details and complex charts, and it can operate across command-line and graphical interfaces. In practice, that means the model is meant to look at a screen, read a chart, and then take actions in software, rather than only answering in text.

Tencent also says the full release substantially scales both pre-training and post-training compared with the preview, with gains in instruction following, long-context understanding, long-horizon tasks and deep research. Those are the skills that matter for agent work, where a model has to carry a task across many steps without losing the thread.

The preview gives a sense of the scale. It had 770 billion total parameters with 49 billion active at a time, and a 1M-token context window. Tencent identified excessive reasoning and over-verification as known issues in that early version, which is the kind of honest caveat that tells you where the preview was rough around the edges.

Pricing stays put, and a cheaper Flash version is coming

Listed API pricing is unchanged from the preview: $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached input tokens. Keeping prices flat while adding vision and computer use is a deliberate move. It tells developers that the upgrade is not a price hike in disguise, and it keeps Hy4 competitive against rival frontier models that charge more for similar agent workloads.

A separate Hy4 Flash release is planned for the near future, aimed at significantly better cost efficiency and faster inference. That is the version to watch if you care about running agents at scale, where the cost per task matters more than raw benchmark scores. Fast, cheap models tend to get used a lot more than expensive, brilliant ones.

Why this release fits a bigger pattern

Hy4 arrives in a week crowded with Chinese model releases built for agents. Meituan's LongCat-2.5 landed on November 2 with native image understanding and a 1.6-trillion-parameter design aimed at long agent tasks. The pattern is consistent: the race has moved from chatbots that answer questions to models that do work across tools, documents and screens.

There is also a compute story here. LongCat-2.5 was trained on domestic Chinese accelerators, and the broader Chinese ecosystem is building around local hardware and local platforms. Hy4's hosted availability through Tencent Cloud fits that picture: the model is a product of an ecosystem, not just a research artifact.

What to watch next

Three things will decide whether Hy4 matters beyond launch week:

• Whether the promised weights and system card actually ship, and what the safety evaluations in that card show.
• Whether computer use works reliably on real desktops, not just demos, where misclicks and ambiguous screens are the norm.
• Whether Hy4 Flash delivers on the cost promise, because agent workloads live or die on price per task.

For now, the sensible take is measured. Hy4's full release is a real upgrade with a clear use case, agent work across vision and software, at unchanged prices. It is not yet the open release many developers will want, and the benchmarks so far come from Tencent itself. Try it on a real task you already understand, compare the output against a model you trust, and judge it on whether it finishes the job without hand-holding. That is the test every agent model eventually has to pass, including this one, and it is the test that OpenAI's GPT-5.6 Sol preview will also face when more details emerge.