Tencent says its new AI model, Hy4, scored 78.9 on OSWorld 2.1, a test of how well an AI can use a computer. In the same chart, Claude Opus 5.5 scored 81.8. But nobody outside Tencent can check these scores, because Hy4's weights are not released yet.
What Tencent Released
On November 4, 2026, Tencent launched the full version of Hy4. It first showed a preview on August 28. Tencent says the new version can look at images and charts. It also says Hy4 can use a computer, both by typing commands and by clicking on-screen apps. Tencent says it trained the model on tools like Blender and CAD software.
Hy4 is a mixture-of-experts model. Think of a big team of specialists where only a few are called in for each job. The preview had 770 billion parameters, which are the numbers a model learns during training. It used only 49 billion for each token, a small piece of text. Tencent has not shared these numbers for the full Hy4.
Open Weights Are Still Only a Promise
Some headlines call Hy4 open-weight. For the full model, that is not true yet. Open weights means anyone can download the model and run it themselves. Only the preview is on Hugging Face, a site for sharing AI models, under the Apache 2.0 license. The preview cannot see images or use a computer.
Tencent says the full weights are coming soon. It has not given a date, a license or a size. Right now, you can only use Hy4 online, through Tencent Cloud, apps like WorkBuddy and CodeBuddy, and OpenRouter. Tencent's TokenHub service is reported to be half price until December 31.
What the Scores Do and Don't Show
OSWorld checks whether an AI can finish tasks on a real computer desktop. Tencent's chart also shows Terminal Bench 4.0, a test of command-line work. There, Hy4 scores 44.4 and Claude Opus 5.5 scores 66.4. Tencent admits it still has a big gap to the best closed models on its hardest tests.
All these scores come from Tencent, and some tests were built by Tencent. Nobody else can repeat them without the weights. In my view, this is why open weights matter: they let others test the claims.
Running Hy4 will not be easy either. All 770 billion parameters must fit in memory, even though only 49 billion work at a time. The preview's setup guide uses eight GPUs, even with a smaller 8-bit version.
What to Watch Next
Tencent says it will release the weights, a report with safety results called a system card, and a cheaper model called Hy4 Flash. The real test is whether outside teams get the same OSWorld score once the weights are out. Anything beyond that is a guess.
Disclosure: This article was drafted with Claude, an AI model made by Anthropic. Tencent's chart compares Hy4 with Claude Opus 5.5.
