Tencent's Hy4 goes live with vision and computer use, weights on hold

Tencent flipped the switch on the full Hy4 model on November 4, 2026. The release adds vision and computer-use abilities, and it's shipping now through hosted APIs and Tencent's own products. The model weights and a full system card? Those are promised later — this launch is hosted-only.

You can get Hy4 through Tencent Cloud and OpenRouter, and it's baked into CodeBuddy, WorkBuddy, Yuanbao and ima. The listed API pricing hasn't moved from the preview: $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input tokens. OpenRouter's deployment listing gives it a 1,048,576-token context window and a 65,536-token completion limit, taking text and images in and returning text — which puts vision inside the hosted API, not just Tencent's own apps. Tencent Cloud is the only provider listed there.

Tencent says the model reads fine-grained image detail and complex charts, and can drive command-line and graphical interfaces. Training aimed at Blender, CAD and long multi-app workflows tied visual perception to planning and tool use. On the coding side, Tencent says frontend work spans tricky UI interactions, business logic, state management and full-stack integration, and its Interactive Entertainment Group pitched in on game-dev data covering Unreal Engine, Unity and Godot, gameplay logic, asset pipelines and multiplayer builds. Tencent reports gains of 19.1% on GameDev and 9.8% on GameLogic over the preview.

Finance got its own grind: Tencent built thousands of realistic workspaces and tens of thousands of tasks, then ran Hy4 past 213 finance professionals — 153 of them from financial institutions — across 579 real-world tasks. In that company-run evaluation, Tencent says Hy4 took first among open-weight models with a score of 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. It also reports hallucinations in internal search evaluations down to 5.45% from 15.76% in the preview, and speculative errors down to 5.6% from 23.1%. On TenureTrail, an internal set of 123 research tasks across seven disciplines, Tencent says Hy4 led open-weight models again.

The August 2026 preview was a different animal — open-sourced on arrival with 770 billion total parameters, 49 billion active and a 1M-token context window, and saddled with known issues around excessive reasoning and over-verification. The full version scales pre- and post-training hard, with Tencent claiming gains in instruction following, long-context understanding, long-horizon tasks and deep research. Still, Tencent admits a gap to top proprietary models, especially on internal benchmarks for research-level math, long-horizon coding, self-improvement and advanced visual question answering.