Tencent released Hy4 on November 4, the latest generation of its large language models. The company states that this release marks an across-the-board capability leap over Hy4 preview, ranking Hy4 among the world's most capable open-weight models for the first time. Tencent plans to release the model weights shortly and publish a system card.
The Hy4 model significantly expands data and training scale across pre- and post-training. It introduces vision capabilities for the first time, which Tencent says delivers improvements in instruction following, long-context understanding, and long-horizon agentic capabilities. These new vision capabilities combine visual understanding, reasoning, and tool use, allowing Hy4 to interpret dense charts, parse documents, and analyze information across multiple images. The model can also combine visual information for multi-step analysis and deduction.
Building on these visual capabilities, Hy4 now supports Computer Use for the first time. It can work across both the command line and graphical interfaces to handle everyday tasks and work. Tencent's Hy team conducted targeted training on specialized software like CAD applications and Blender, as well as long-horizon tasks spanning multiple applications to improve the model's ability to plan and execute professional workflows.
Tencent states that the Hy team works with product teams on "Co-Design" within real-world enterprise workflows, specifically focusing on WorkBuddy, Tencent's AI agent for office work. In office scenarios, Hy4 improved the WorkBuddy user experience, achieving an average quality score of 83.18 and a pass rate of 96.15% in a WorkBuddy-focused evaluation. Since its March 2026 launch, WorkBuddy has released more than 100 major version updates, seen a 50-fold growth in token consumption, and completed more than 9 billion task requests. WorkBuddy is currently the largest AI agent for office productivity in China by user base, with its enterprise edition deployed across more than 50 industries, serving 50% of China's top 500 enterprises and 60% of its top 100 enterprises.
Hy4 also provides a major upgrade in professional financial analysis. In a blind evaluation by 213 finance experts—153 of whom were from institutions outside Tencent and roughly seven in ten had more than five years of experience—Hy4 ranked first in overall performance among the open-weight models evaluated across 579 real-world finance tasks.
In front-end and game development, the Hy team made targeted optimizations for complex interactions, business logic, and state management to ensure generated interfaces are functionally complete. The team also worked with professional game developers at Tencent Interactive Entertainment Group (IEG) to improve capabilities such as game feature implementation and asset processing. Hy4 is adapted to engines like Unreal Engine, Unity, and Godot. Compared with Hy4 preview, Hy4 improved by 19.1% on GameDev evaluations and 9.8% on GameLogic evaluations.
For scientific research, Hy4 combines complex reasoning and scientific programming. In TenureTrail, an internal evaluation of 123 executable research tasks across seven disciplines, Hy4 demonstrated capabilities in areas including mathematical construction, physical modeling, and socio-economic decision-making.
To address risks in cybersecurity and biology, the Hy team took a responsible approach during training. On cybersecurity, Hy4 was trained only on defensive data; offensive material was excluded to avoid any potential uplift in offensive capabilities. In-house experts from Tencent Xuanwu Lab and Keen Lab audited the model, finding no notable risk in its offensive capabilities. Regarding biological safety, Hy4's open-ended experiment design lags frontier models by a wide margin (30.4/100). Internal experts found a solid grasp of foundational biology but limited ability to turn that into executable experimental plans, with no evidence that Hy4 can meaningfully lower the barrier to carrying out biochemical harm.
Hy4 is now live across Tencent's product portfolio, with its API available through third-party platforms including OpenRouter, Hermes, and OpenCode. Tencent is keeping Hy4 preview's pricing—$0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million tokens on cache hits—despite the capability uplift. Hy4 flash will be released shortly to offer a lower cost per task. From November 4 to November 15, WorkBuddy and CodeBuddy will offer a limited quota of free usage. Additionally, from today through December 31, 2026, Hy4 will be billed at 50% of its list price on Tencent Cloud TokenHub.
Tencent Hy is preparing for future research into recursive self-improvement (RSI). Since rebuilding its AI infrastructure in February 2026, the company has shipped a major model roughly every two months, scaling from under 300 billion parameters in Hy3 to 770 billion in Hy4. The Tencent Hy team stated that while Hy4 is among the most capable open-weight models today, a considerable gap remains between it and leading closed-source models, and they will continue to scale training to close that gap.
