Tencent Launches Hy4, Its First AI Model That Can See and Operate a Computer
Tencent Launches Hy4, Its First AI Model That Can See and Operate a Computer
Tencent released the full version of Hy4 on November 4, 2026, adding two capabilities that are new to its Hunyuan model family: vision and computer use. The model can now interpret images and charts, and it can work through command-line and graphical interfaces to carry out multi-step professional tasks. It is available now through hosted APIs and inside Tencent products, although the open model weights Tencent promised have not been released yet.
From a text-only preview to a multimodal flagship
Hy4 is the successor to the Hy4 preview Tencent released in August 2026. That preview was a text-only model with 770 billion total parameters, 49 billion of them active in a mixture-of-experts design, and a context window of one million tokens. Tencent says the full release scales both pre-training and post-training, with gains over the preview in instruction following, long-context understanding and long-horizon tasks.
The headline change is vision. For the first time in the Hunyuan series, a model can look at images and reason about what it sees. Tencent says Hy4 can interpret fine-grained image details and complex charts, parse documents and analyze information across multiple images. In a post on X, the Tencent AI account summed it up in one line: “Introducing Hy4. It can see now.”
The second new capability is computer use. Hy4 can operate across command-line and graphical interfaces to complete professional and everyday tasks. Tencent gave it targeted training on specialized software such as Blender and CAD applications, along with long-horizon tasks spanning multiple applications.
Built alongside the products that use it
Tencent developed Hy4 in close collaboration with the product teams that deploy it, an approach it calls Co-Design. The main partner is WorkBuddy, its office productivity agent. In a WorkBuddy-focused evaluation, Hy4 achieved an average quality score of 83.18 and a pass rate of 96.15 percent.
Financial analysis received the same treatment. Tencent built thousands of realistic workspaces and more than 10,000 tasks covering fund research, securities analysis and quantitative strategy work, then evaluated the model with 213 finance professionals, including 153 experts from financial institutions outside Tencent, across 579 real-world tasks. In that company-run evaluation, Hy4 ranked first among the open-weight models tested, with a score of 3.04, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each.
Game development was another focus area. Developers from Tencent Interactive Entertainment Group helped curate training data and evaluations covering Unreal Engine, Unity and Godot. Tencent reports that Hy4 improved by 19.1 percent on its GameDev evaluation and 9.8 percent on GameLogic compared with the preview.
Access, pricing and the missing weights
Hy4 is available now across Tencent’s product portfolio, including WorkBuddy, CodeBuddy, Yuanbao and ima, with API access through Tencent Cloud and third-party platforms including OpenRouter, Hermes and OpenCode. OpenRouter’s listing specifies a 1,048,576-token context window and a 65,536-token completion limit.
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. WorkBuddy and CodeBuddy offer a limited quota of free usage from November 4 to November 15, and Hy4 is billed at 50 percent of its list price on Tencent Cloud TokenHub through December 31, 2026. A cheaper and faster variant, Hy4 Flash, is planned to follow shortly.
The main caveat is the open-weight release. Tencent says it plans to publish the model weights and a comprehensive system card in the coming days, with support developed alongside the vLLM and SGLang communities. As of the launch, only the August preview weights were publicly available on Hugging Face under the Apache 2.0 license. Until the full release weights arrive, Hy4 can only be used through hosted APIs.
Safety work and an honest gap
Tencent says it took a more cautious approach to safety in this release. Cybersecurity training used only defensive data related to proof-of-concept generation and vulnerability patching, while material involving exploitation and penetration testing was left out. Experts from Tencent Xuanwu Lab and Keen Lab audited the model and reported no elevated risk.
On biological safety, Hy4 scored 30.4 out of 100 on Hy-BioExperimentBench, an internal benchmark for experiment design. Internal reviewers found its ability to turn biological knowledge into executable experimental plans remains limited, with no evidence that it meaningfully lowers the barrier to biological harm.
Tencent is also candid about where Hy4 stands. The company says the model has joined the ranks of the strongest open-weight models for the first time, while acknowledging a considerable gap to the leading closed-source models, particularly on research-level mathematics and long-horizon coding. Model scale grew from under 300 billion parameters in Hy3 to 770 billion in Hy4, with Hy5 already on the horizon.
Why it matters
Hy4 matters for two reasons. First, it brings vision and computer use, the capabilities that turn a chatbot into an agent that can work inside real software, to a model family Tencent intends to open up. If the weights arrive as promised, developers will be able to run a 770-billion-parameter model that can see screens and operate applications on their own infrastructure. Second, the launch shows how quickly the open-weight field is moving: Tencent has released a major model roughly every two months since rebuilding its AI infrastructure in February 2026. The test now is whether the promised weights and system card follow quickly, because until they do, Hy4 is an open-weight model in name only.
