Tencent Hunyuan Hy4, released on 4 November 2026, is the first Hunyuan model that can see as well as reason. It processes images, analyses complex charts, and does visual mathematical reasoning, combined with reasoning and tool calls. It also adds Computer Use support, so it can operate a computer through graphical interfaces and the command line. For a business owner, that combination is the point: an AI that can read a chart, open the right application, and carry out a multi-step task beats a chatbot that only answers questions.

This is the general release of Hy4, following the Hy4 preview in August 2026. Tencent says it joins the strongest open-weight models, while trailing leading closed models on research maths, long-horizon coding, and advanced visual question answering. Weights are to be open-sourced soon, at announcement. This post explains what Hy4 can do, where it could help, and how to pilot it safely.

Tencent Hunyuan Hy4: Visual AI That Can Operate Your Business Software

What Hy4 actually is, and what it is not

Hy4 is a mixture-of-experts model with 770 billion total parameters, about 2.6 times Hy3's 295 billion. It combines language reasoning, visual understanding, and tool use.

Two firsts matter most. Visual understanding is new to the Hunyuan line. Computer Use is also new: Hy4 can operate computers via graphical interfaces and command lines, with targeted training for Blender and CAD and for cross-application, long-horizon tasks — built to work across applications, not just inside a chat window.

What it is not, yet, is a finished, fully open product you can download and run anywhere. Tencent says the weights will be open-sourced soon, and a lighter, cheaper companion, Hy4 Flash, is announced as coming soon, not yet released. Treat it as a newly released model with a clear direction, not a settled platform.

Visual understanding: reading charts, documents, and screens

Most business information does not arrive as clean text: it is a scanned invoice, a dashboard screenshot, a spreadsheet chart, or a photo of a whiteboard. An AI that cannot read those formats forces your team to retype everything first.

Hy4's visual capability targets that gap. It can extract figures from a chart, follow a diagram, and combine what it sees with what it reads — relevant to finance and reporting.

Where visual reading helps first

• Finance and reporting: reading charts and dashboards, then summarising trends and flagging anomalies for a human to check.
• Operations: interpreting forms, labels, and photos from the shop floor or warehouse.
• Design and engineering: working with drawings alongside text specifications.

Human checking remains essential: a misread scale makes a wrong finance figure worse than no figure. Use assisted reading — the model drafts, a person verifies. That guarded workflow is central to good AI agent development and AI automation.

Computer Use: an AI that operates software, not just talks about it

Computer Use may change how businesses think about AI agents. Instead of answering a question about a task, Hy4 is designed to carry out the task in software: opening applications, using their interfaces, running commands, and moving between applications. Tencent highlights targeted training for Blender and CAD and cross-application, long-horizon tasks.

That appeals to front-end and game development, where work spans design tools, code editors, and test environments. It also matters for ordinary business software: copying data between systems or preparing a report from several sources. Repetitive work can be drafted by the model and reviewed by a person.

What to expect, realistically

• Long tasks need supervision. One wrong step early on can derail the rest.
• Access control is essential. An agent that operates software must be limited to the applications, data, and actions it needs.
• Start with repeatable tasks. A well-defined, repeated workflow is a far better pilot than a vague, one-off job.

See our guide to AI agents for small business automation and AI chatbot vs AI agent. The lesson: an agent's permissions and approval steps matter as much as its capability.

Where Hy4 could help a real business

Hy4 is integrated into Tencent products including WorkBuddy, CodeBuddy, Yuanbao, ima, and the WeCom assistant Dayuan, and is available via Tencent Cloud TokenHub. That shows where Tencent sees the first value: workplace assistance, coding help, and document work. For other businesses, the question is which of your workflows match what Hy4 is built for.

Four areas stand out. First, finance analysis: reading reports and charts, reconciling figures, and preparing first drafts for review. Tencent reports a financial crowd-test with Tencent Financial Technology (213 experts, 579 real tasks) in which Tencent says Hy4 ranks first among open-source models. That is Tencent's claim, not an independently verified result — a signal about finance capability, not proof it will handle your accounts unaided.

Second, software work: front-end and game development benefit from a model that can see a design and work in a development tool. Third, document-heavy operations: reading and acting on documents mixing text, tables, and images. Fourth, customer support, where a screenshot can make the difference between a useful answer and a generic one. Our chatbot development and WhatsApp bot development work deals with that daily: customers send photos the bot must understand.

Open-weight ambition, with honest limits

Tencent positions Hy4 among the strongest open-weight models, while trailing leading closed models on research maths, long-horizon coding, and advanced visual question answering. That honesty is useful: Hy4 looks strong for practical work, not a breakthrough that makes specialist tools obsolete.

The open-weight direction matters for control: businesses could run, adapt, and inspect a model under their own governance, rather than depending entirely on a hosted service. That fits our open-weights guide for business and our work on governed AI agents: capability is half the decision; deployment and oversight are the other half.

There is also a safety note. Tencent says it filtered exploit-development, reverse-engineering, and penetration-testing content from training to limit offensive hacking capability. That is a design choice, not a guarantee. Deploying a computer-using agent still needs access controls, logging, and human approval for sensitive actions.

Finally, do not plan around Hy4 Flash yet. It is announced as coming soon, but not released. If cost is your main concern, wait for its actual release terms.

A pilot plan that does not bet the business

Test Hy4 with a small, contained pilot: one workflow, one team, one month, and a clear measure of success. Decide in advance what "good" looks like. A pilot that tries to automate everything at once teaches nothing.

A practical four-step pilot

1. Pick a visual, repeatable task. For example, summarising weekly sales charts.
2. Keep a human in the loop. Every output gets checked before it reaches a customer or a live system.
3. Limit access from day one. Give the agent read-only or sandboxed access first, and only the applications it needs.
4. Measure time saved and errors caught. If it saves an hour a week but introduces one serious error, it has failed.

That discipline applies to custom software, a SaaS product, a mobile app through Flutter app development, or an online store via ecommerce development. The model is the engine; workflow, permissions, and checks make it safe. To scope a pilot, see our contact page and about page.

The same logic applies in our core markets: United Kingdom, UAE, Saudi Arabia, Pakistan, and USA.

Frequently asked questions

What is Tencent Hunyuan Hy4?

Hy4 is Tencent's Hunyuan model released on 4 November 2026, following the August 2026 preview. It combines reasoning and tool calls with visual understanding and Computer Use support, with 770 billion total parameters. See the SCMP report on the Hy4 unveiling.

Is Hy4 open source yet?

Not at announcement. Tencent says the weights are to be open-sourced soon, and Hy4 is available through Tencent products and Tencent Cloud TokenHub. Until then, treat open-weight availability as announced, not delivered.

Can Hy4 really operate business software?

That is the aim of Computer Use: operating computers through graphical interfaces and command lines, with targeted training for Blender and CAD and cross-application tasks. Expect supervised, well-scoped use first, with limited permissions and human review.

Should my business switch to Hy4 now?

Probably not as a wholesale switch. Run a small pilot on one repeatable, visual workflow, measure time saved and errors caught, and keep a human in the loop. Tencent's broader enterprise AI direction is outlined on its press page on Tencent Cloud enterprise AI.