Tencent releases Hy4 with vision and computer use, with model weights still to come

Tencent released Hy4 on November 4, 2026, adding visual understanding and the ability to operate software to its latest language model, according to the company. Access is available through Tencent products, with third-party developer access rolling out. Tencent calls Hy4 open-source, but says model weights and system cards will follow in the coming days.

From reading images to operating software

The release follows Hy4 Preview, which Tencent released and open-sourced on August 28. That earlier model targeted coding, office work and scientific research. November’s announcement adds visual understanding and computer use for the first time, extending the model’s stated capabilities beyond the preview’s productivity focus.

Tencent says Hy4 can analyze detailed images, interpret complex charts and reason through mathematical problems presented visually. Its computer-use capabilities cover both command-line interfaces, where instructions are typed, and graphical interfaces, where software is operated through on-screen controls. The company also trained it specifically on professional software, including Blender and CAD tools.

Inside WorkBuddy, the emphasis is on connecting those abilities to information already used at work. Through integration with WeCom, Tencent says Hy4 can draw on chats, email, documents, calendars and meetings. It can track project history, identify unfinished tasks and flag meetings or deadlines that need attention. Team collaboration through WorkBuddy’s planned Projects feature remains an upcoming addition.

The reliability gains are measured internally

Tencent links its workplace ambitions to targeted search improvements: retrieving current information, combining multiple sources and checking facts for consistency. In internal evaluations, it reports fewer hallucinations—incorrect or invented information—and fewer errors caused by unsupported assumptions than with Hy4 Preview.

Tencent’s internal error measurements

Hallucination rate

12.95%→7.2%

%

Tencent reports a decline in hallucination rate from Hy4 Preview to Hy4 in internal evaluations.

Errors from unsupported assumptions

23.1%→5.6%

%

Tencent reports a decline in errors caused by unsupported assumptions in the same internal comparison.

For financial analysis, Tencent describes a blind evaluation conducted by its Hy team with Tencent Financial Technology. It involved 213 financial professionals assessing 579 real-world tasks. Of those participants, 153 came from external institutions, and approximately 70% of the full group had more than five years of industry experience.

Hy4 received an overall score of 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97. Kimi K3.1 Pro and GLM-5.3 each scored 2.68. Those results are Tencent’s reported comparison, not an independent ranking. The company says its financial training drew on thousands of specialized workspaces and tens of thousands of tasks covering fund research, securities analysis and quantitative investment.

Product access now, downloadable files later

Tencent lists Hy4 as available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom’s AI Assistant and ima. It says access through an API—a connection developers use to call the model from their own software—is being made available through OpenRouter, Hermes and OpenCode. Integrations with QQ, Honor of Kings and other Tencent products are still underway.

Published pricing and launch offers

• Standard pricing: approximately S$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens.
• WorkBuddy and CodeBuddy: free access from November 4 through November 15.
• Tencent Cloud TokenHub: Hy4 at 50% off standard pricing through December 31, 2026.

Ahead of the promised file release, Tencent describes safeguards focused on cybersecurity and biological safety. It says training prioritized defensive security work and excluded offensive exploitation materials. Assessments by Tencent’s Xuanwu and Keen security labs found no significant concerns about offensive capabilities in the areas evaluated. Internal biological reviews found strong foundational knowledge but more limited open-ended experimental design than frontier models.