Platforms

Tencent's Hy4 Brings 770B-Parameter Multimodal Reasoning and Native Computer Use to Spatial and Industrial AI

The full Hunyuan Hy4 release adds visual understanding and native computer-use to a 770-billion-parameter frontier model now accessible on OpenRouter and embedded in Tencent's enterprise productivity suite.

Mainstay Communications Team · November 6, 2026 · 4 min read

Tencent's Hy4 Brings 770B-Parameter Multimodal Reasoning and Native Computer Use to Spatial and Industrial AI

In brief

Tencent shipped the full Hy4 on November 4, 2026, upgrading the model from a text-only preview to a multimodal system that can analyze images, interpret charts, and operate desktop software through both command-line and graphical interfaces. The 770-billion-parameter mixture-of-experts model carries a 1-million-token context window and is live on OpenRouter, Tencent Cloud TokenHub, and embedded in WorkBuddy and CodeBuddy. Its training dataset includes Blender, CAD applications, Unreal Engine, and over 100 sandboxed professional tools, making it the first frontier-class model in Tencent's lineup to target agentic workflows in design, engineering, and 3D production environments.

The Release

Tencent shipped the full Hy4 on November 4, 2026, eight weeks after releasing the text-only Hy4 preview on August 28. The preview opened as an open-weight model under Apache 2.0 and established a commercial API baseline. The November release layered on what the company had withheld: native visual understanding, computer-use operation, and a 50% pricing discount through December 31 on Tencent Cloud TokenHub.

Dowson Tong, Tencent's senior executive vice president, telegraphed the direction at the company's Global Digital Ecosystem Conference on October 29. "Tencent is developing its models and products together to improve their performance in practical applications," Tong said. Full weights and a system card are planned for release; they weren't available at launch.

Architecture and Scale

Hy4 runs a mixture-of-experts design with 770 billion total parameters, of which 49 billion are active during each inference pass. The architecture uses 78 layers (1 dense, 77 MoE) and routes across 256 specialized experts plus 1 shared expert, activating 8 per token. The context window is 1 million tokens, 4 times the 256,000-token ceiling of its predecessor Hy3.

The scale increase from Hy3's 295 billion total parameters to 770 billion is the largest jump in the Hunyuan series. Tencent says it used high-quality training data co-created with internal specialists in software engineering, gaming, finance, and security, alongside product teams from WorkBuddy and CodeBuddy.

Computer Use and Visual Reasoning

The two headline additions in the November release are fine-grained visual understanding and native computer operation. For vision, Tencent's evaluations highlight chart analysis, mathematical diagram reasoning, and high-resolution image processing. The company demonstrated the model locating a roof fitting roughly 2 meters across inside a 2.8-billion-pixel aerial photograph.

For computer use, Hy4 can navigate graphical interfaces, execute command-line tasks, and chain multi-step workflows across applications. Its training corpus included Blender, CAD tools, Unreal Engine, Unity, Godot, and more than 100 sandboxed professional applications. That specificity positions the model for spatial and 3D production workflows rather than generic web automation.

Platform Footprint

Hy4 is available through OpenRouter under the model ID tencent/hy4, with API pricing at $0.417 per million input tokens and $1.251 per million output tokens. Direct access through Tencent Cloud TokenHub runs $0.834 per million input and $2.501 per million output, discounted 50% through December 31. Cached tokens cost $0.042 per million.

The model is embedded in WorkBuddy, Tencent's productivity agent, which reached 32 million monthly active users and is used by roughly half of China's top 500 companies. CodeBuddy, the company's AI coding assistant, also runs Hy4 as a primary model option. Additional access points include Yuanbao, ima, and WeCom's Dayuan assistant.

WorkBuddy handled over 9 billion task requests between its March 2026 launch and the Hy4 full release. The platform connects to enterprise messaging systems and offers computer-use workflows that read calendars, documents, emails, and meetings.

Tencent's predecessor model, Hy3, saw weekly usage climb 68 times over between its preview and final versions. Hy4's commercial strategy follows the same pattern: open-source the preview weights early to seed the developer ecosystem, then ship the full capability set through integrated products.

Performance Benchmarks

On vendor-reported evaluations, Hy4 scored 78.9 on OSWorld 2.1, a benchmark for desktop computer use, behind Claude at 81.8 and ahead of GPT-6.1 Sol at 75.2. On E-Bench 2.0, the model scored 70.0 against Claude's 80.0. Terminal-Bench 4.0 placed Hy4 at 44.4, with Claude Opus 5.5 at 66.4 and GPT-6.1 Sol at 55.5.

The preview version posted separate benchmark results before the visual and computer-use updates shipped. On DeepSWE, a software-engineering benchmark, it scored 64.3, ahead of DeepSeek-V4 Pro at 62.7 and Alibaba's Qwen-3.8 Max at 56.6. Code Arena WebDev ranked the preview 8th globally. A blind engineering evaluation across 203 tasks run by 163 experts gave the model 2.99 out of 4.00, compared to Kimi K3 at 2.94 and GLM-5.3 at 2.92. All benchmark figures are vendor-reported and have not been independently reproduced.

What Comes Next

Tencent confirmed one more release in the pipeline. Hy4 Flash will carry the same visual understanding and agent capabilities at lower cost and higher throughput, targeting latency-sensitive enterprise deployments.

The open-weight release for the full Hy4 is pending. Apache 2.0 licensing is confirmed, consistent with the preview and with Hy3 before it. Weights will ship with a technical system card once Tencent publishes them.

Sources

Companies mentioned

Mainstay Digital · Spatial and 3D visualization for industrial companies

Back to Newsroom