On November 4, 2026, Tencent announced the full release of its AI model Hy4. Compared with the preview version, it expands the scale of pre-training and post-training, and for the first time supports visual understanding and PC control. Tencent says it can not only read images and charts, but also operate software through command-line interfaces (CLIs) and graphical user interfaces (GUIs). One example of an intended use would be reading a chart in a document and using that information to carry out work in another application.
However, to use such capabilities in real work, what matters is not only whether answers are accurate but also whether the model can reliably complete the operations it is told to perform. Tencent has presented performance improvements from internal evaluations and a training policy aimed at limiting the risk of misuse, but its official materials contain inconsistencies in how they describe the evaluation results and the training content. It requires careful scrutiny to determine what the published figures actually measured and how far they support the reliability of PC control.
From Image Understanding to PC Control
According to the development team's explanation, Hy4 combines visual understanding, reasoning, and tool use, and handles image processing, complex chart comprehension, and mathematical reasoning that draws on visual information. For PC control, it handles both CLIs and GUIs, and the team says it was also trained on tasks spanning multiple applications, including the 3D creation software Blender and CAD software.
Explaining the contents of a chart mainly tests whether the model can read the information correctly and answer. By contrast, when it uses that information to operate software, it must also be evaluated on whether it completed the task from start to finish as instructed. The task examples shown in public demos alone are not enough to judge the success rate on tasks in typical business settings.
Did the Error Rate Improve? Discrepancies Remain in the Official Materials
The official announcement and the development team's explanation, both published on November 4, agree on one figure: "errors from guessing" fell from 23.1% to 5.6%. On the wrong-answer rate, however, the numbers differ: the official announcement gives 12.95% to 7.2%, while the development team's explanation gives 15.76% to 5.45%. The evaluation targets also differ, with search optimization in one and finance and office work in the other.
| Official source | Wrong-answer rate (preview → full release) | Errors from guessing (preview → full release) | Evaluation scope described |
|---|---|---|---|
| Official announcement | 12.95% → 7.2% | 23.1% → 5.6% | Internal evaluation on search optimization |
| Development team's explanation | 15.76% → 5.45% | 23.1% → 5.6% | Internal evaluation on finance and office work |
Both materials report performance improvements, but neither discloses how the evaluation targets were selected, how many tests were run, how the metrics were defined, or how results were aggregated. It is therefore impossible to tell whether the difference in figures stems from differing evaluation conditions or from an inconsistency in the write-ups. There is also no basis for averaging the two values or adopting one as the correct figure.
The development team says performance improved in areas such as searching for information that changes over time, integrating multiple sources, and checking the consistency of facts. "Errors from guessing" here refers to errors based on unsupported assumptions, a metric separate from the wrong-answer rate. Neither indicates a failure rate for PC control, and better search or answer accuracy does not necessarily mean a higher success rate on work spanning multiple applications.
In game development evaluations as well, Tencent reports improvements over the preview version of 19.1% on "GameDev" and 9.8% on "GameLogic." However, as SiliconANGLE points out, the baseline scores used to calculate these improvement rates have not been published. These are improvement rates reported by Tencent, not independently measured results from an outside organization.
Training Policy to Limit Offensive Capabilities, and Safety Evaluations
A model that can operate a PC also carries the risk of being misused for cyberattacks. Tencent has described a training policy that emphasizes defensive uses, but the official materials do not agree on what was included in the training data.
The development team's explanation says cybersecurity training was limited to data for defensive purposes, such as generating proofs of concept (PoCs) that verify a vulnerability exists and fixing vulnerabilities. It says that not only attack-code development but also reverse engineering, which examines how software works, and penetration testing were excluded. The official announcement, on the other hand, says reverse engineering and penetration testing for analytical purposes were incorporated into training, while material that could be exploited for attacks was excluded. It is unclear what criteria were used to classify the training data or how the two accounts can be reconciled.
On safety, Tencent says its subsidiaries Xuanwu Lab and Keen Security Lab examined tasks involving cryptanalysis, reverse engineering, and vulnerability exploitation and found no increase in risk. However, this was an evaluation conducted by Tencent itself, and it does not constitute independent verification or a guarantee of safety.
In biology, the development team reported a score of 30.4/100 on "Hy-BioExperimentBench" and says this fell below the high-risk threshold it had set for itself. The specific threshold value has not been published. Internal experts assessed that the model's ability to turn knowledge into executable experimental procedures is limited. Training for defensive uses, assessing the risk of misuse for attacks, and operating software accurately in everyday work are separate matters, and the success rate of PC control cannot be judged from these safety evaluations.
Release of Model Weights and the System Card Is the Next Focus
Hy4 is available through Tencent Cloud's TokenHub and OpenRouter, and is also offered in WorkBuddy, CodeBuddy, Yuanbao, and ima. According to Tencent's announcement, API pricing is unchanged from the preview version, WorkBuddy and CodeBuddy are free to use from November 4 to 15, and TokenHub plans to offer it at 50% of standard rates until December 31, 2026. A lightweight version, "Hy4 Flash," is also said to be coming soon.
Meanwhile, as of the November 4 announcement, the full release's model weights and the system card summarizing safety evaluations and other details were still only planned for release. The official announcement says they will be published within the next few days, and the development team also mentions collaboration with vLLM and SGLang. Being able to use the model through an API is not the same as being able to verify the model weights or evaluation conditions externally, and the license terms for the full release cannot be determined from the materials available at the time of the announcement.
The development team also acknowledges that on research-level math problems and long-running coding tasks, a large performance gap remains with the top proprietary models. The same applies to the model's ability to improve itself and to advanced image question answering. These added capabilities do not guarantee that it can complete complex work without human intervention.
To judge Hy4's practical usefulness, verification is essential that clarifies not only the evaluation conditions and metric definitions behind the error rates but also what counts as task success in PC control. Key points to watch next are whether the discrepancies in the descriptions of cybersecurity training are resolved, and whether the release of the model weights and system card allows outside parties to reproduce the evaluations.