Tencent Hy4: API Pricing, Vision, Computer Use and Open-Weight Plans

Tencent released Hy4 on November 4, adding visual understanding and computer use to its flagship model. The API is available through Tencent Cloud and OpenRouter, with a million-token context window and unchanged list pricing. Downloadable weights are still planned, rather than part of today's release.

The useful change is what vision allows an agent to check. A coding model can write a chart or a user interface from text instructions. A model that can also inspect the result has another way to catch mistakes before handing the work back.

What changed from Hy4 preview

In its Hy4 announcement, Tencent describes larger pre-training and post-training runs, improvements to instruction following and long-context understanding, and new visual capabilities. It specifically calls out work across command-line and graphical interfaces, including training on Blender, CAD and workflows spanning multiple applications.

For developers, this makes Hy4 relevant to more than generating code. An agent could inspect a screenshot, change the application, then inspect it again. That still requires a harness which supplies screenshots, executes tools and controls permissions. Buying access to the model does not by itself provide a safe, complete computer-use product.

Tencent also reports better game development and office-document work. Its examples cover game engines such as Unity, Unreal Engine and Godot, and editing Word, Excel and PowerPoint files. These are company demonstrations and evaluations, not evidence that every generated project is production-ready.

Hy4 API pricing and context window

Tencent's announced list prices are unchanged from Hy4 preview. At publication, OpenRouter's Hy4 listing shows a separate 50% discount through its Tencent Cloud provider.

Per million tokens	Tencent announced list price	OpenRouter displayed discounted price
Input	$0.834	$0.417
Output	$2.501	$1.251
Cached input	$0.042	$0.021

These are US-dollar token prices, not a fixed price for completing a task. A long-running agent can consume several calls, repeated screenshots and tool results. Budgeting from the price of one prompt would miss most of that work. The displayed discount should also be checked again before forecasting ongoing costs.

OpenRouter lists a 1,048,576-token context window and up to 65,536 completion tokens. The model accepts text and images and returns text. Its API identifier is tencent/hy4.

The listing currently has one provider, Tencent Cloud, and says EU in-region routing is unavailable. Teams with regional processing requirements should resolve that before integrating it. A million-token context limit does not remove deployment or data-handling constraints.

How to read the Hy4 benchmarks

Tencent's announcement contains several useful before-and-after comparisons, but their scope matters. In its internal evaluations described in its finance and office-workflow section, the reported hallucination rate falls from 15.76% for Hy4 preview to 5.45%. The company also describes assessments involving 213 finance professionals across 579 tasks.

Those results support a narrower conclusion: Tencent measured an improvement on the work and evaluation setup it selected. They do not establish a universal error rate for financial advice, spreadsheets or an application's own documents. Different retrieval sources, tool permissions and instructions can change the result substantially.

The strongest practical test is a completed artifact. For a spreadsheet agent, that means checking formulas, source references and totals after the file has been edited. For a coding agent, it means running the application and testing the interaction, rather than accepting a convincing explanation of what the code should do.

Tencent itself acknowledges remaining gaps against top proprietary models on difficult mathematics, long-horizon coding, model self-improvement and advanced visual question answering. That is a more useful limitation than treating one headline score as a ranking for every workflow.

Are Hy4 weights available?

The API release and the weight release are separate. Tencent says it plans to release the weights with a comprehensive system card and is working with the vLLM and SGLang communities on support. Its announcement does not give a completed download release.

OpenRouter already displays a model-weights link, but that destination returned a 404 when checked for this article. A link in a catalogue is not proof that the checkpoint is downloadable. Local deployment requirements, licensing and the full system card should be assessed when the actual release is available.

Hy4 is also available in Tencent products including CodeBuddy, WorkBuddy, Yuanbao and ima, according to the announcement. Some accompanying team-collaboration features in WorkBuddy are described as forthcoming. The separately planned Hy4 flash model should not be confused with the model released today.

What developers should test first

The sensible starting point is one visual workflow with a clear success check: a broken interface, a chart that needs verification, or a document whose formatting and calculations can be inspected. Compare the finished result, retries, latency and total token bill with the model already doing that job.

If Hy4's visual feedback reduces the number of corrections a person has to make, unchanged token pricing can still produce a cheaper workflow. That is the result worth measuring while the wider open-weight release remains ahead.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn launches and product decisions into clear stories that developers can understand and share.

Book a 30-minute call.