Tencent released Hy4, an open-source model that can use a computer

A team picking an open-weight model for an agent that runs a CAD tool had no option trained on CAD. Tencent said on 4 November that its new Hy4 model was trained on Blender and CAD. Computer use is built in.

Hy4 is the latest Hy large language model from Tencent. The company said it will release the weights and system cards in the coming days. The post said safety focus was on cybersecurity and biological safety. The release followed the Hy4 Preview from an earlier date.

What the model can do, in Tencent's words

Hy4 adds vision input for the first time. Tencent said this covers image understanding, chart analysis and mathematical visual reasoning. Hy4 also adds computer use. The model can act in command-line interfaces and in graphical user interfaces. It can run work and everyday tasks that need both.

The model received specialised training on professional software such as Blender and CAD, Tencent said. The company said this improved long-horizon planning on complex tasks. Hy4 keeps improvements in instruction following and long-context handling, Tencent wrote.

The benchmark scores Tencent shared

Tencent said Hy4 is the top-performing open-source model on a blind evaluation of financial tasks it ran. The test used 213 financial professionals. 153 of them were external experts. About 70 percent had more than five years of industry experience. The test ran 579 real-world financial tasks.

On that test, Tencent published four scores. Hy4 scored 3.04. DeepSeek-V4.1-Flash-1017 scored 2.97. Kimi K3.1 Pro scored 2.68. GLM-5.3 scored 2.68. All four scores are Tencent's own figures. The company ran the test with Tencent Financial Technology, its own unit.

On Tencent's internal WorkBuddy evaluation, Hy4 scored 83.18 on average. The pass rate was 96.15 percent. Tencent said Hy4 ranked first in nine of twelve scenarios in the mainland Chinese market. These are all Tencent's own numbers.

Tencent also said Hy4 cut its hallucination rate from 12.95 percent in the Preview version to 7.2 percent. Errors from unsupported assumptions fell from 23.1 percent to 5.6 percent. These are internal evaluations Tencent ran itself.

Game and science work

On game development, Tencent said Hy4 improved by 19.1 percent on its GameDev test. It improved by 9.8 percent on GameLogic. Both figures compare to the Preview version. The model supports Unreal Engine, Unity and Godot. Tencent said it worked with Tencent Interactive Entertainment Group on this training. All these results come from Tencent's own internal tests.

For science work, Tencent said it tested Hy4 on 123 tasks across seven disciplines. The test was called TenureTrail. The company claims Hy4 did well on mathematical construction, physical modelling and systems engineering. Tencent also open-sourced two foundational models trained with Hy4. HyCell covers single-cell representations. HyPDE covers physics fields.

What a buyer should check

Every figure above is Tencent's. None of the benchmarks are from an outside party. The news that matters is a model trained on professional software and vision, with weights to come. The news that does not yet matter is where Hy4 lands on tests the buyer did not pick.

Three facts the post does not settle. The licence that covers the weights. Whether computer use works outside the mainland Chinese market tools Tencent tested it on. And the model size, which the post does not give. Wait for the system card.

Source

Tencent, "Tencent Launches Hy4, Joining the Ranks of the World's Leading Open-Source AI Models".