What's New in Tencent Hy4? Vision, Computer Use, Benchmarks

What is new in Tencent's Hy4?

Hy4 is Tencent's follow-up to the text-only preview it released at the end of August. The main changes are vision input and computer use, meaning the model can work in a command line and also click through a graphical interface. Tencent also trained it on tasks like Blender and CAD. It reports gains over the preview on its benchmarks. The catch: the preview shipped as open weights under Apache, while Hy4 is API only for now.

How does Hy4 differ from the preview?

The preview was a text model. You gave it text and got text back. Hy4 widens that in two directions.

Vision matters mostly because of the second change. A model that can't see a screen can't reliably operate one. Computer use depends on reading screenshots, locating buttons and fields, and then issuing clicks or keystrokes. Pairing that with a command line is practical, since many real workflows mix both: run a build in a terminal, check the result in a browser, adjust a setting in a desktop app.

Why train on Blender and CAD?

Blender and CAD are tool-heavy, visual, and structured. Both reward models that can follow multi-step procedures and check spatial results. Training on them suggests Tencent wants Hy4 to produce and manipulate 3D or design artifacts, not just describe them. The sources do not say how the training data was built or how much of it there was, so treat the motivation as inference. The pairing with the GameDev and Game Logic results is at least consistent with a focus on game and 3D content workflows.

What do the reported benchmarks show?

Tencent's numbers, as relayed in the source:

• GameDev: up about 19% over the preview.
• Game Logic: up nearly 10% over the preview.
• Apex Agents: 75.2, versus 74.8 for the competing model listed as Opus 5.5.
• Training-system optimization: in internal tests, the model found bottlenecks and sped up throughput by about 20%.

The sources give no absolute scores for the GameDev and Game Logic results, only relative improvements. They also do not describe the test setups, so you can't tell how large the base scores were. A 19% gain on a low base means something different from 19% on a high one.

How much should you trust the Apex Agents score?

Treat it as a promising claim, not a settled ranking. Three reasons.

1. It is self-reported. Tencent ran the evaluation. Vendor-run benchmarks often use favorable prompts, tool configurations, or retry budgets.
2. The margin is small. 75.2 against 74.8 is 0.4 points. On agentic benchmarks, run-to-run variance can be larger than that.
3. Setup details are missing. The sources don't say how many runs were averaged or what scaffolding each model used.

None of this means the result is wrong. It means the honest reading is "competitive with a top-tier model on this benchmark," not "better." Independent evaluations, once the model is widely accessible, will tell you more.

What does the 20% throughput claim mean?

Tencent says that in internal tests the model located bottlenecks in training systems and improved throughput by roughly 20%. That is a notable claim about agentic engineering ability, since it implies the model can profile a real system, form a diagnosis, and change code or configuration. It is also the least verifiable item here. It was an internal test on Tencent's own infrastructure, and the sources give no detail on the workload, the baseline, or how much human guidance was involved. Read it as a sign of the use case Tencent is targeting.

Will Hy4's weights be released?

Tencent says yes, soon. For now, Hy4 is available through an API only. That is a real change from the preview, which came with full open weights under an Apache license. Apache is a permissive license that allows commercial use, modification, and redistribution, which is why the preview was attractive for self-hosting and fine-tuning.

"Coming soon" has no date in the sources, and nothing says which license the Hy4 weights will carry. If you plan to build on Hy4 locally, the sensible approach is to wait for the actual release and read the license before committing. If an API is fine for your use case, you can test the model now without depending on that promise.

Who should care about Hy4 right now?

• Agent builders who want a model that can drive both a terminal and a GUI, and who are comparing options for computer-use work.
• Game and 3D tooling teams, given the GameDev, Game Logic, Blender, and CAD emphasis.
• Self-hosters and fine-tuners, who should watch for the weights release and its license terms rather than assume they match the preview.

What is still unknown?

The sources leave out several things you would want before adopting a model:

• Parameter count and architecture
• API pricing and rate limits
• Context window size
• Hardware requirements for local use, once weights arrive
• A release date and license for the weights
• Independent benchmark results

I won't guess at any of these. Check Tencent's official release notes and model documentation for specifics.