Tencent Hy4 Launches With Vision and Computer Use, Weights Next

Update (November 4, 2026): This story now includes more figures from Tencent’s full press release, covering WorkBuddy results, hallucination rates, a new finance benchmark and two open-source science models.

Tencent released Tencent Hy4 on November 4, a new flagship AI model that adds image understanding and computer use for the first time. The model is available now through Tencent Cloud and OpenRouter at $0.834 per million input tokens and $2.501 per million output tokens, and Tencent says it will publish the open weights and a system card in the coming days.

In its Hy4 research post, the Hy team says the release places its model family among the strongest open-weight models for the first time. It also admits that “a notable gap remains” with top proprietary models on research-level math, long-horizon coding and advanced visual question answering.

What Tencent Hy4 Can Do

Hy4 can read fine details in images, analyze complex charts and solve visual math problems. Its new computer use mode works in both command-line tools and graphical interfaces, and Tencent says it trained the model on professional software such as Blender and CAD so it can plan and finish longer multi-step tasks.

For developers, Tencent highlights frontend work and game development. The company says it tuned Hy4 for Unreal Engine, Unity and Godot with game developers from its Interactive Entertainment Group, and reports gains of 19.1% on its GameDev benchmark and 9.8% on GameLogic compared with Hy4 preview, which launched in August.

Benchmarks and Finance Focus

Most of the numbers Tencent shared come from its own tests. In a blind evaluation with 213 finance professionals across 579 real tasks, Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68). Tencent also says errors caused by unsupported assumptions fell from 23.1% to 5.6% after it reworked the model’s search features.

The company did not publish a parameter count in its announcement, but TechNode reports that Hy4 has 770 billion parameters. Independent benchmark results should follow once the weights are public.

What Else Tencent’s Tests Show

The full press release adds more results, all from Tencent’s internal evaluations. On WorkBuddy, the company’s AI work assistant, Hy4 posted an average score of 83.18 and a 96.15% pass rate, and Tencent says it came first in nine of the 12 scenarios it tested in mainland China. Through WorkBuddy, the model also hooks into WeCom to pull context from chats, emails, documents, calendars and meetings.

The same search rework that cut unsupported assumptions also brought Hy4’s hallucination rate down to 7.2%, from 12.95% for Hy4 preview. On the finance side, Hy4 gained 12.03 percentage points over the preview on iRaB, an investment research agent benchmark run by financial AI company Rabyte. None of these figures have been checked independently yet.

Tencent’s Hy team also open-sourced two science models built with Hy4 and its Hyra 2.0 harness: HyCell for single-cell representations and HyPDE for physics fields. Outside Tencent’s own services, developers can reach the Hy4 API through OpenRouter, Hermes and OpenCode.

Price, Availability and Open Weights

Hy4 keeps the same API price as Hy4 preview. Tencent Cloud’s TokenHub is offering it at 50% off through December 31, 2026, and the WorkBuddy and CodeBuddy assistants are free to use from November 4 to 15. Tencent is also bringing the model to Yuanbao, ima and WeCom, with QQ and Tencent News integrations in progress.

On safety, Tencent says it limited cybersecurity training data to defensive material and ran red-team audits with its Xuanwu and Keen security labs before release. A cheaper and faster Hy4 Flash model is planned next, according to the company’s press release.

Why It Matters

The launch adds another strong Chinese contender to an open-weight race that already includes DeepSeek, Moonshot’s Kimi and Z.ai’s GLM, whose cost advantage recently led Revolut to move 2,000 engineers to self-hosted GLM. Open models also matter more for Chinese companies as access to Western compute narrows, as the reportedly canceled Oracle Tencent chip deal showed this week.

For businesses outside China, downloadable weights would let them run Hy4 on their own hardware, a route that releases like Mistral Large 4 have already made popular. Until the weights ship, Hy4 is available only through APIs and Tencent’s own apps.