An AI assistant can write a convincing plan and still struggle to finish the job. A chart may contain a crucial label that never reaches the text prompt. A browser action may change the page in a way the agent fails to notice. Tencent Hy4 targets that gap between understanding a request and completing it across applications.
Tencent announced Hy4 on November 4, 2026, adding visual understanding to its productivity-focused model family. For developers and product teams, the question is whether those capabilities make it easier to complete work across documents, applications, and tools.
What changes after Hy4 preview
The earlier Hy4 preview arrived on August 28. Its official repository describes a mixture-of-experts model with a 770-billion-parameter backbone, 49 billion parameters active per token, and a one-million-token context window. It is released under Apache 2.0. Those are the preview’s specifications and license terms; Tencent’s November announcement does not establish that they apply to the new release.
According to Tencent’s release announcement, Hy4 supports computer use through both command-line and graphical interfaces, with targeted training for Blender and CAD workflows. Hy4 is available through Tencent Cloud and OpenRouter, and in CodeBuddy, WorkBuddy, Yuanbao, and ima. Tencent plans to release the model weights and a fuller system card later. Hy4 Flash is also planned for a later release.
That leaves two different adoption paths. Teams can evaluate the hosted model now. Teams considering a self-hosted deployment still need the weights, license terms, and hardware requirements to assess whether it is practical. The preview’s deployment instructions do not establish what the new release will require.
Why vision matters for agents
Vision gives an agent another way to inspect the state of its work. A developer could ask it to compare a rendered page with a reference screenshot, then investigate a misplaced control. An office assistant could examine a chart alongside the spreadsheet that produced it. These examples suggest useful tests; they are not results from hands-on testing.
Imagine asking an agent to repair a form whose submit button disappears on a narrow screen. Reading source code may suggest a CSS change. Looking at the rendered page can reveal whether the button is actually visible, whether a fixed footer covers it, and whether the change breaks another viewport. The useful result is a working form, with evidence that the repair held up.
OpenRouter lists Hy4 as accepting text and images and producing text. To act inside an application, the model needs an agent framework that provides tool access and executes its requested actions. Image understanding alone does not grant access to a browser, filesystem, or business account.
The preview repository also acknowledges excessive reasoning time and over-verification as limitations. That history makes elapsed time an important part of an upgrade test. A model that inspects more screens can still leave a user waiting longer for the same completed task.
What the evaluations establish
Tencent reports a finance evaluation involving 213 finance professionals and 579 real-world tasks. Hy4 averaged 3.04 out of 4, compared with 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for both Kimi K3.1 Pro and GLM-5.3. These are Tencent-organized results. Its release also acknowledges a remaining gap with leading proprietary models on difficult internal tests.
The finance scores offer a reason to run a closer comparison for teams producing research documents or spreadsheet models. Their relevance depends on how closely the evaluation resembles a team’s own work. Tool access, available data, and the definition of a successful result all affect an agent evaluation.
For a useful trial, keep the starting files and instructions identical across models. Require an output that can be checked: a spreadsheet with working formulas, a report with traceable sources, or an application whose requested features work. Record where a person had to intervene. That helps show whether better model outputs translate into less manual work.
Pricing depends on where you access Hy4
Tencent’s English release lists the same standard API rates as Hy4 preview. OpenRouter displayed a 50% discount on those rates when checked on November 4, 2026. All figures below are US dollars per million tokens.
Token category | Tencent release rate | OpenRouter discounted rate |
|---|---|---|
Uncached input | $0.834 | $0.417 |
Cached input | $0.042 | $0.021 |
Output | $2.501 | $1.251 |
The OpenRouter model page identifies the model as tencent/hy4, with a 1,048,576-token context window and up to 65,536 completion tokens. It also lists tool calling and structured outputs. These discounted prices were listed on November 4, 2026; check the current rates before estimating a production budget.
At Tencent’s standard rates, 100,000 uncached input tokens and 10,000 output tokens cost about $0.1084. That calculation covers token charges only. An agent that repeats a large prompt across several calls, retries failed actions, or uses separately billed tools will have a different total.
Compare the cost of getting a usable result as well as the price per token. A cheaper run that needs substantial manual repair may consume more of a team’s budget once review time is included. Conversely, repeated tasks with reusable context may benefit from caching, provided requests actually receive cache hits.
Tencent Hy4 deserves a trial where visual context and sustained tool use are central to the task. Start with a job your team already understands well enough to judge, and measure completion time, corrections, and total cost alongside output quality. Those results will show whether the new capabilities earn a place in your workflow.