Tencent Hy4 Brings Vision to Everyday AI Work
Explore How Tencent Hy4 Connects Visual Understanding With Coding, Office Work, and Research. Review the Official Benchmark Gains, Compare Input and Output Costs, and Understand the Difference Between Available API Access and the Planned Model-Weight Release Before Choosing What to Test.
Tencent Hy4 adds visual understanding to a model built for work that spans documents, software, and multiple tools. Released on November 4, 2026, it expands beyond text-only assistance toward interpreting images, navigating interfaces, and carrying longer tasks through to a useful result. Tencent also says its API pricing is unchanged from Hy4 preview.
The practical question is how much these changes help your own workflow. This review examines Tencent's published capabilities, evaluation results, and launch pricing. The evidence comes from the official Hy4 announcement, rather than an independent hands-on test.
What Is New in Tencent Hy4?
Tencent describes larger-scale pre-training and post-training than Hy4 preview, alongside improvements in instruction following, long-context understanding, and deep research. In plain language, the goal is a model that follows detailed requirements and keeps track of more information across a task.
Vision is the headline addition. Hy4 can interpret image details and complex charts, connecting what it sees with reasoning and tool use. It also introduces computer use across command-line interfaces, where operations are typed, and graphical interfaces, where applications expose visible controls.
These capabilities matter together. Reading a chart is one task; using its information to update a report or guide an application workflow is a more demanding one. Tencent reports targeted training for software including Blender and CAD, plus workflows that cross applications. That makes visual productivity a useful evaluation angle, without assuming every application is supported equally well.
Where the New Capabilities Could Help
Frontend and Game Development
Tencent's development overview highlights interface interactions, business logic, state management, and full-stack integration. For creators, this shifts the evaluation beyond attractive screenshots: a generated interface should also preserve inputs, handle navigation, and behave correctly when users change settings.
For games, Tencent describes work on gameplay logic, state tracking, and asset pipelines, with adaptation for Unreal Engine, Unity, and Godot. A sensible trial would use one small, testable feature, such as inventory persistence or a menu interaction. Check the resulting behavior in the engine before expanding the task.
For the visual side of a prototype, SuperMaker's AI Image Maker offers text-to-image and image-to-image tools for exploring scene concepts and supporting artwork. Keep that asset-creation step separate from testing Hy4's code and gameplay logic.
Documents and Research Workflows
The announcement describes improvements in search, multi-source synthesis, and factual consistency checks, alongside collaborative editing of Word, Excel, and PowerPoint files. These are relevant to reports that need both readable explanations and traceable supporting data.
For a content team, a useful evaluation might combine a source document, a chart, and a reporting template. Ask the model to produce a concise brief while preserving units, dates, and source references. Measure how much correction the output needs, rather than judging fluency alone. This is a suggested test, not a result demonstrated here.
Once your brief is checked, you can turn its creative direction into a short visual explainer with SuperMaker's AI Video Maker. Specify the subject, action, and camera movement, then review the generated clip against your brief before using it in a presentation.
Tencent Hy4 Benchmarks: Read the Gains in Context
Tencent publishes several comparisons with Hy4 preview. The following figures come from its finance and office evaluation discussion and development section.
| Evaluation | Published Result | Relevant Context |
|---|---|---|
| GameDev | 19.1% improvement over Hy4 preview | Tencent-reported game development evaluation |
| GameLogic | 9.8% improvement over Hy4 preview | Separate gameplay-logic evaluation |
| Hallucination rate | 15.76% to 5.45% | Internal evaluation described in the finance and office section |
| Speculative errors | 23.1% to 5.6% | Internal evaluation, distinct from hallucination rate |
The error reductions are particularly relevant to document-heavy work, where a plausible but unsupported statement can be costly to correct. However, these percentages describe Tencent's evaluation conditions. They are not a universal error rate for every prompt or product integration.
Do not combine the rows into one overall performance score. Game development, gameplay logic, and factual reliability measure different things. The useful next step is to reproduce your own representative tasks with fixed inputs and a clear acceptance checklist.
Tencent Hy4 Pricing and Access
As of November 4, 2026, Tencent's release and pricing section lists these API rates:
| Token Type | Price per Million Tokens |
|---|---|
| Cached input | $0.042 |
| Input | $0.834 |
| Output | $2.501 |
Tokens are the units used to meter model input and output. Cached input is priced separately, so do not apply the lowest rate to all requests. As a simple calculation, one million ordinary input tokens plus one million output tokens would cost $3.335 at the listed rates, before any additional service charges.
Tencent says Hy4 is available through Tencent Cloud and OpenRouter, and in products including CodeBuddy, WorkBuddy, Yuanbao, and ima. The launch page also links to Hy AI Studio. Check the chosen platform's current billing and access conditions before estimating a production workload; the announcement's token prices do not establish every product's subscription terms.
Is Hy4 Available for Self-Hosting?
The announcement uses open-model language, but its release section explicitly says Tencent plans to open-source Hy4's weights. It also says a comprehensive system card will accompany that release, with ecosystem support involving vLLM and SGLang communities.
For deployment planning, distinguish available hosted access from the planned weight release. The announcement alone does not establish a downloadable checkpoint, a finalized license, or hardware requirements for your self-hosted installation. Teams needing those details should wait for the actual release artifacts before committing infrastructure.
Should You Test Tencent Hy4?
Tencent Hy4 is worth evaluating if your work combines visual material with coding, document editing, or tool-driven research. Start with one bounded workflow and compare completion quality, correction time, and total token cost against your current setup.
Keep expectations grounded. In its limitations discussion, Tencent acknowledges gaps behind leading proprietary models in research mathematics, long-horizon coding, model self-improvement, and advanced visual question answering. Hy4 flash is described as a future release, not the model covered here.
The strongest reason to test Hy4 is the combination of new visual capabilities and unchanged launch API pricing. Its value for your team will depend on reliable task completion, not the breadth of the announcement alone.
