Tencent launches Hy4, a 770-billion-parameter open-source model with vision and computer-use skills

What is Hy4?

Hy4 is Tencent's latest-generation large language model, succeeding the Hy4 preview that shipped in August 2026 without vision or computer-use features. The full release adds native multimodal capabilities and agentic functionality that positions it as a direct competitor to models like Claude, GPT-4, and Gemini.

The model uses a 770-billion-parameter mixture-of-experts architecture, activating only a subset of parameters per inference to balance compute efficiency with output quality. It supports a 1-million-token context window, enabling it to process book-length documents, extensive codebases, and multi-hour meeting transcripts in a single prompt.

What new capabilities does Hy4 add?

For the first time in the Hunyuan line, Hy4 ships with native visual understanding integrated into the core model rather than as a separate vision encoder. It can analyze images, diagrams, screenshots, and charts, then answer questions or generate text based on visual content. This enables use cases ranging from technical documentation generation from architectural diagrams to visual question-answering in customer support.

Hy4 also introduces native computer-use functionality, allowing the model to interact with graphical user interfaces, execute commands, and control software applications. The system can navigate web browsers, fill forms, operate desktop applications, and automate multi-step workflows. Early demonstrations show Hy4 completing tasks like booking appointments through web interfaces and managing email workflows inside Tencent's WorkBuddy and WeChat Work platforms.

The model has undergone specialized training on professional software including Blender and CAD tools, significantly enhancing its ability to plan and execute long-horizon, sophisticated tasks. This positions Hy4 as a direct competitor to Anthropic's Claude computer-use capabilities and Google's emerging agent frameworks.

How does Hy4 perform on benchmarks?

Tencent reports significant improvements over the preview version across multiple dimensions. In a blind evaluation conducted with 213 financial professionals, including 153 experts from leading external institutions, Hy4 achieved the highest overall performance score among open-source models tested on 579 real-world financial tasks.

The model also improved its hallucination rate from 12.95% in Hy4 Preview to 7.2%, while errors caused by unsupported assumptions fell from 23.1% to 5.6%. In the Investment Research Agent Benchmark (iRaB), Hy4 achieved a 12.03 percentage-point improvement over the preview.

On game development benchmarks, Hy4 improved by 19.1% on GameDev and 9.8% on GameLogic compared to the preview version, reflecting specialized training on Unreal Engine, Unity, and Godot integrations.

How can developers access Hy4?

Hy4 is available across Tencent's product ecosystem including Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, and the WeChat Work intelligent assistant Dayuan. Third-party API providers including OpenRouter, Hermes, and OpenCode integrated Hy4 on launch day.

Despite significant performance improvements over Hy4 Preview, Tencent maintains the same pricing: approximately $0.90 per million input tokens, $2.70 per million output tokens, and $0.045 per million cached input tokens. Through 31 December 2026, Tencent is offering a 50% discount on TokenHub API pricing and free access to WorkBuddy and CodeBuddy from 4-15 November.

The open-weights release under Apache 2.0 allows developers to self-host Hy4 for on-premises deployments, addressing data-residency and compliance requirements in regulated industries.

What does this mean for the open-source AI landscape?

Tencent's Hy4 launch signals accelerating progress in Chinese open-source AI development, particularly in multimodal and agentic capabilities. By combining 770B-scale language modeling with vision and computer-use skills in a single open architecture, Tencent delivers a credible alternative to closed Western models for developers building multimodal workflows or automation agents.Earlier this year, DeepSeek and MiniMax cleared China's new AI export registration for open-source releases, suggesting a broader push for Chinese open-source AI to compete globally.

For enterprises evaluating open-source LLM options, Hy4's computer-operation capabilities and 1M-token context make it a strong candidate for workflow automation and document-processing pipelines, especially in markets where data sovereignty favors on-premises deployments over cloud API calls.