Tencent launches Hy4 with visual understanding, computer use and a million-token context window, extending its flagship model beyond text into image analysis and software operation. The November 4 release is already available through Tencent products and third-party API platforms.

 

The company calls Hy4 open-source, but the launch arrived before its model weights and system card. Tencent says both will follow in the coming days, leaving researchers temporarily unable to inspect the downloadable model or its fuller safety evidence.

 

The Hy4 release introduces four notable changes:

  • Image understanding and visual reasoning
  • Command-line and graphical computer use
  • A 1,048,576-token context window
  • Defensive limits on cybersecurity training

 

Background Reading

 

Tencent Launches Hy4 With Vision and Computer Use

Hy4 is Tencent's first model in the series to accept images. The company says it can handle fine-grained image processing, complex chart analysis and mathematical visual reasoning, while maintaining the instruction-following and long-context capabilities expected from a general-purpose language model.

 

Computer use broadens that scope. According to Tencent's release announcement, Hy4 can work through command-line and graphical interfaces. Specialized training covered professional applications including Blender and computer-aided design software, targeting tasks that require planning across many steps.

 

Those capabilities put Hy4 in competition with models designed to operate tools rather than only answer prompts. The practical question is whether it can reliably preserve state, recover from mistakes and complete long workflows inside real software. Tencent supplied internal benchmark claims, but independent evaluations were not available at launch.

 

OpenRouter lists a 1,048,576-token context window and up to 65,536 output tokens. It says the hosted model accepts text and images, supports function calling and structured JSON output, and was available from Tencent Cloud on release day.

 

Hy4 Hosted Access Arrives Before Model Weights

Tencent made Hy4 immediately accessible through Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's AI Assistant and ima. API access is also being distributed through OpenRouter, Hermes and OpenCode, while Tencent plans integrations with QQ, Honor of Kings, Game of Peace, WeRead and Tencent News.

 

The company prices standard access at about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens. A TokenHub promotion cuts that pricing by half through December 31, while WorkBuddy and CodeBuddy are free from November 4 through November 15.

 

Hosted availability does not provide the same transparency as downloadable weights. Tencent's announcement says the weights and system cards will arrive in the coming days. Until those files appear, developers cannot verify the promised open release terms, run the model on their own infrastructure or inspect the detailed evaluation methodology.

 

A lower-cost Hy4 Flash model is also scheduled to follow. Tencent has not yet published enough detail to compare its architecture, speed or quality with the main release, so buyers should treat the Flash variant as a roadmap item rather than a current option.

 

Tencent Limits Offensive Cybersecurity Training

Tencent says Hy4's cybersecurity training emphasized defensive work such as vulnerability identification, remediation and analytical proof-of-concept development. It excluded offensive exploitation material intended to build hacking tools or break into systems, according to the company and Reuters.

 

Internal specialists from Tencent's Xuanwu Lab and Keen Security Lab evaluated cryptographic reverse engineering, binary deobfuscation and vulnerability exploitation. Tencent reported no significant offensive-capability concerns in the tested areas. That conclusion is company testing, however, and the promised system card will be necessary to judge coverage, thresholds and limitations.

 

The company applied a similar framing to biological risk. Its internal reviewers found strong foundational biology knowledge but more limited performance on open-ended experimental design than unnamed frontier models. Tencent has not yet released the full test set or independent validation behind that comparison.

 

Hy4 Claims Still Need Independent Testing

Hy4 combines three trends in current model development: multimodal input, tool-driven agents and very long context. Together they could make the model useful for document-heavy analysis, visual tasks and software workflows, especially inside Tencent's large product ecosystem.

 

The release also illustrates why publication timing matters. A hosted endpoint lets users test outputs, but model weights and a system card enable deeper scrutiny. Tencent's open-source description will become easier to assess only after those artifacts appear with clear licenses, reproducible evaluations and sufficient technical detail.

 

Independent testers should focus on computer-use reliability, prompt injection, data leakage, long-context retrieval and the boundary between defensive security assistance and actionable exploitation. Hy4's launch is substantial, but its strongest safety and openness claims remain provisional until the promised documentation and weights are public.

 

Illustrative photograph: Tencent headquarters campus in Shenzhen by Lhzss8 via Wikimedia Commons, licensed under CC BY-SA 4.0. Cropped to 16:9; the adapted image is shared under the same license.