Concept illustration of visual information transformed into continuous action through the model, and left behind to be checked

The concept of the cover is a concept generated by AI, not the actual architecture, experimental equipment or performance results of Hy4.

Interpretation date: 2026-11-04
Deadline: 2026-11-04 20:37:34 (Beijing time)

On November 4, Tencent announced Hy4, saying it had added visual understanding and computer use. It said the weights for the official release and the system card would be available in the coming days. For researchers, the question is whether combining visual input with multistep action yields an improvement they can test. A release announcement, access to a service, and model materials available for independent study each provide different evidence. Tencent issued a statement

The significance of the new ability is that the action chain has grown longer

Tencent stated that Hy4 can process visual information such as images and charts, and operate the command line and graphical interface; the team has also trained for professional software such as Blender and CAD. These are the descriptions of the developer's ability, and this article has not been measured. The announcement date is today, and the specific training and evaluation time has not been disclosed, and the release date cannot be regarded as the completion date of the experiment. ability statement

Computer operations can be commonly understood as: the model not only answers "how to do", but also reads the interface, selects actions, and then continues to make decisions according to the feedback after the operation. The research value of this direction is to put perception and execution into the same observable process.

For example, reading a scientific research chart and correctly sorting the data into the file is two different levels of tasks. The former may only need to be judged once, and the latter also involves positioning content, selection tools, processing format and check output. A recognition error may be enlarged in subsequent steps or may be corrected by feedback. Therefore, the single picture Q&A score cannot directly replace the reliability measurement of the complete task.

The preview version has materials and cannot be automatically transferred to the official version

Tencent's official website records the release of Hy4 preview in August 28. It is the historical background of this official release, not the new opening result today. Preview Announcement

The currently-readable preview repository lists the main parameter 770b, 49b for each tokens, and it is clear that the parameter table does not contain additional multi-tokens prediction layers. The so-called mixed expert model is to select only some experts to participate in each calculation. This caliber reminds us that even if we just copy the amount of parameters, we need to indicate the version and statistical scope. Preview repository

The corresponding page on Hugging Face is still explicitly named Hy4-preview, labeled as a text generation model, and lists the Apache-2.0 license. It supports us to discuss the status of the preview version of the material status, but it is not enough to prove that the official version uses the same license, the exact same structure, or the same deployment conditions. Preview model card

As of the deadline of this article, the official copyright weight and system card that can be checked have not been obtained in this search; the announcement said that it will be provided later. This report is the range of materials that have been verified, and does not infer the entire release status with one page access failure. The existing repository, model card and official website are mainly from the same developer, and cannot be counted as three independent verifications.

Experts participate in the evaluation, and it is still necessary to identify who has organized the comparison

Tencent issued a statement to introduce a blind comment jointly conducted by the Hunyuan team and Tencent Fintech: 213 professionals participated, of which 153 were from external institutions, covering 579 financial tasks. This article retains its organizational relationship and sample size, and does not directly write the participation of professionals as an independent reproduction. Evaluation description

Blind reviews usually reduce the first impression of the brand by hiding the model identity. However, it will not automatically solve the problems of how to sample tasks, how many people are scored by each question, and whether the tool permissions are the same. The 213 review does not mean that each task has obtained 213 independent judgments, and it cannot multiply the number of people with the number of tasks as a known total.

For researchers, the scoring rules, task distribution, and per-question results would be more revealing. Experts might favor an answer because it is correct, clearly presented, or complete in its procedure. Those dimensions need separate assessment. To tell whether a difference is stable, we also need to know how much scores vary and how failures count. I could not obtain the per-question records needed to recalculate the comparison, so it provides no basis here for a definitive model ranking.

The value of the system card is to put the ability back into the conditions

The system card is a document describing the model evaluation, usage conditions and limitations. For long-tasking agents, the reader needs to know which model version is being measured, and what tools, tips and operating environments are paired with.

This is the follow-up inspection direction proposed in this article: the same model is changed to a set of tools, or more opportunities to retry, and the final completion rate may change. The complete comparison should be described with the peripheral executing program, and explain time-consuming, number of calls and manual intervention. Otherwise, it is difficult to judge whether the progress comes from the model itself or the process configuration.

What is especially worth saving is the failure trajectory. When the task is not completed, is it not clear the interface, the plan is wrong, or is it executed correctly but there is no verification file? Many mechanisms cannot be distinguished if there is only a final screenshot. The process of public playback can often help researchers locate the improvement space than adding a few more successful demonstrations.

Next, wait for the evidence that can change the judgment

This release explains the development direction, and also puts forward the proposition of verifiable ability. At this stage, the evidence mainly comes from institutional announcements and existing preview materials, which are not peer review studies that have been reviewed in this article; this search has not found an independent reproduction of these official version propositions, and this article does not run a model.

The next step is to observe whether the official version of the material can correspond to a clear version, and whether the external team can reproduce the results under the same budget and the same budget. You should also see whether the system can still complete the work when the interface changes, task interruptions and error feedback appear.

If these evidences are gradually supplemented, the meaning of Hy4 can be advanced from "declaring joining vision and operation" to "what tasks and which conditions are more reliable". Today, we can seriously pay attention to this research route, and at the same time leave the conclusion where the evidence has reached.

References and reading range

  1. Tencent:Tencent Launches Hy4, 2026-11-04. The original statement issued by Tencent through PR News Agency; the direct opening failed, and the text of the original site that has been read by the retrieval tool has been read, including paragraphs such as ability, evaluation, and open plan. The page displays 19:13, the time zone of the display time is not independently verified, and the exact release time is not calculated accordingly.
  2. Tencent:Tencent Releases and Open-Sources Tencent Hy4 preview, 2026-08-28. The official website announcement has been read, which is used to check the historical version and date.
  3. Tencent-Hunyuan/Hy4-preview, continue to update the page, the current version update time has not been verified. The structure, parameter caliber, evaluation and limitation paragraph of the read README; no weights or code was downloaded.
  4. tencent/Hy4-preview model card, continue to update the page, the current version update time has not been verified. Read model labels, license tags, and model descriptions; do not use them as official versions of the system card.

Apps by the author

The following is the author's own product promotion. In addition to writing a technical blog, I have also independently developed two apps to help organize daily life. Welcome to understand:

  • Daily Expense Journal: Record daily income and expenditure, arrange budget and savings goals, and make each account and plan more clear. View on the App Store.
  • One Small Action: Choose small actions according to the current time and energy, record the feelings after completion, and leave a clue to the daily review. View on the App Store.

Find more articles and projects on my website: zffjzz.top. Your feedback helps me improve these tools.