Tencent Releases Hy4 Official Version: First Introduction of Visual Capabilities, Hands-On Testing of Operating WeChat and 3D Modeling

On November 4, Zhidongxi reported that today Tencent released the official version of its latest-generation flagship model, Hy4, and plans to open-source the model weights. Compared with the Hy4 preview released in August this year, Hy4 significantly expanded the scale of pre-training and post-training, introduced visual capabilities for the first time, and substantially improved comprehensive capabilities such as instruction following, long-document understanding, long-horizon tasks, and deep research.

The addition of visual capabilities from scratch unlocked many new ways to use the Hy model. For example, users can upload casually taken photos and have Hy4 determine the geographic location and verify the information. Or, based on an image of a textbook page, create a 3D explanatory webpage.

According to the complete benchmark results published by Tencent, Hy4 surpassed models such as DeepSeek-V4.1-Flash on multiple public benchmarks in areas including agentic coding, search, and office work, but there is still a certain gap between it and models such as GPT-6.1 Sol and Claude Opus 5.5.

Hy4 is also preparing for recursive self-improvement (RSI). After combining with Tencent's self-developed scientific research agent Hyra2.0, Hy4 optimized the pre-training infrastructure for internal small models of the same architecture, achieving a 20.1% end-to-end training throughput improvement, and automatically trained the Jev model, which outperforms the official one on multiple evaluations.

Hy4 is already available on Tencent Cloud and OpenRouter, and can be experienced in Tencent AI products such as CodeBuddy/WorkBuddy, Yuanbao, and ima. While its performance significantly exceeds that of Hy4 preview, it maintains the same API price. Hy4's API price falls between the input and output prices of DeepSeek V4 Pro during idle and peak periods, while the cache-hit price is consistent with the peak-period price of DeepSeek V4 Pro.

Tencent also plans to release Hy4 flash in the near future, which will significantly improve the model's cost-effectiveness and inference speed.

Tencent said it has joined forces with the vLLM/SGLang community and Inferact Inc. to provide open-source ecosystem support for Hy4. Before the release, the domestic computing power ecosystem had already begun inference adaptation for Hy4. On November 2, the vLLM-Ascend open-source plugin repository for Ascend NPUs added a PR titled "Add Hy4 deployment guide," covering single-node, multi-node, and PD-disaggregated deployment of Hy4 on Ascend 950DT.

Before the official release of Hy4, Zhidongxi obtained early access in advance and conducted hands-on testing. From our test results, adding native visual understanding is the most substantive change in Hy4; it directly changes the interaction mode for tasks such as coding and 3D modeling, and the experience of operating PC applications such as WeChat for Windows is also better. Equally noteworthy is the improvement in long-horizon task capability: across multiple rounds of iteration, Hy4 showed stronger goal retention than previous Hy-series models.

1. Native visual understanding makes its debut; image-based coding and 3D modeling pass hands-on tests

Native visual understanding capability had previously been a blank spot for Hy flagship models. In this hands-on test, we examined it together with complex tasks such as coding and 3D modeling.

In coding scenarios, Tencent carried out targeted optimization for complex front-end interactions, business logic, state management, and front-end/back-end linkage, so that the interfaces generated by the model are not only good-looking but also have complete and usable functionality.

First, let's look at a lightweight webpage redesign task. I sent the link to Zhidongxi's official website to Hy4 and selected another website as a reference design, asking Hy4 to redesign it accordingly.

After receiving the task, Hy4 fully explored both websites. On the one hand, it read the HTML code to become familiar with the website's structural skeleton and style genes; on the other hand, through screenshots, it confirmed design details that are difficult to perceive directly from code, such as layout proportions, visual focus, and color temperament.

During the exploration, it also scraped all the image assets from the Zhidongxi official website to facilitate subsequent development.

Native visual understanding also made code modification much easier: I could directly take screenshots and point out to the model the places on the interface I wanted to change, without having to laboriously describe the layout, color scheme, and element positions in text. For small modifications, HY4 takes roughly only 1-2 minutes, and its chain of thought appears more concise than that of Hy4 Preview.

Ultimately, after three or four rounds of human feedback, Hy4 delivered a redesign of the Zhidongxi official website with which we were relatively satisfied. While retaining details such as Zhidongxi's brand color and website interactions, it adopted TechCrunch's layout and avoided typical AI-flavored design language such as particle light effects and gradient glows.

Our next hands-on task was 3D model development, in the Blender environment. Tencent said that on Hy4, they conducted specialized training for professional software such as Blender and CAD and for cross-application long-horizon tasks, optimizing the planning and execution of professional tasks.

To make good use of external tools, the corresponding Skills and connected MCP connectors are indispensable. I found a collection of Blender Skills on GitHub and asked Hy4 to complete the installation and MCP configuration.

After completion, I uploaded a 2D image of Tencent's classic penguin character and asked Hy4 to build a 3D model based on it, with a reserved embedding spot for a screen on the penguin's belly, so that it could become the shell of an AI hardware device.

After 17 minutes of execution, Hy4 delivered the first version of the model. This model captured the key features—the red scarf, round belly, and screen embedding spot were all there—but overall fidelity still had room for improvement.

At the same time, it also fully considered other elements required for AI hardware, and without any explicit request from us, added a USB-C opening, speaker openings, a microphone opening on top of the head, and internal assembly reference parts.

Over the next 30-plus minutes, I took screenshots while conversing with Hy4, going through 6 rounds of revisions, pointing out one by one the differences between its model design and Tencent's classic penguin character.

The problem that repeatedly failed to be fixed was the design of the penguin's left eye. In the reference image, the eye is a white vertical oval, but in the first several versions Hy4 made it into a black arc floating on the face; only after it enlarged the reference image 4x and 8x did it see clearly and correct it.

However, during the revision process, the 3D models Hy4 delivered did not suffer from the problem of "drifting further with each revision"; each version showed perceptible improvement. The final product it delivered is as follows. Compared with the first version of the model, its fidelity is noticeably better, and at the same time engineering usability was not affected by the appearance rework.

Looking at the whole process, Hy4 did a good job of geometric self-checking in 3D modeling; its weakness lies mainly in visual self-checking. This may be related to the parametric modeling approach currently adopted by most large models for 3D modeling: what the parametric pipeline can directly read back are all scalars, and when facing cross-object, cross-scale problems, it is prone to gaps in judgment.

From this round of experience, what is worth noting is Hy4's performance on long-horizon tasks. Whether it was the multiple rounds of feedback on the webpage redesign or the 3D modeling task that consumed a 500,000-token context window, Hy4 remembered the original task goal. For Agent-type products that require long-chain execution, this capability is crucial.

2. Letting Hy4 be a worker for a day: making a financial report PPT, scrolling through WeChat, and not fabricating even when blocked

In addition to the above complex long-horizon tasks, we also examined Hy4's performance on office-type tasks. The task prompt was as follows:

Make full use of the data and materials from the Q3 financial report on the SpaceX IR webpage, including the Earnings Update, Q&A, 10-Q, and Transcript, to generate a PPT for presenting business progress.

After receiving the task, Hy4 first found the website and downloaded the files mentioned in the prompt. Among them, the PDF file of the Earnings Update downloaded extremely slowly, so it chose a workaround to obtain the same-source full text from the SEC official website, using the original PDF only for archival verification.

When downloading the Q&A file, Hy4 encountered a regional block. It did not fabricate content; instead, it supplemented information through retrieval and noted this limitation in its answer.

After about 30 minutes, Hy4 completed the generation of the first version of the PPT file. We spot-checked several data points, and they all matched the real situation. However, its first version of the PPT was all densely packed text, with a poor reading experience.

After I pointed out this problem, Hy4 used a PDF screenshot tool to save the rich visual materials in the SpaceX financial report locally and incorporated them into the PPT design, ultimately producing a PPT that combines text and images, balancing aesthetics and information density.

With native visual capabilities, Hy4's computer operation abilities are also worth trying. Today, the process of letting Hy4 operate WeChat and automatically post comments in the comment section can already run through completely. Its operation positioning is precise, but execution speed is still relatively slow.

▲The execution process of Hy4 posting a comment (20x speed)

The problem of grabbing the mouse also objectively exists. When it uses the computer, I cannot perform other operations, otherwise it may interrupt its normal execution.

I also tried letting Hy4 scroll through a WeChat group and summarize work messages. It took about 2 minutes to complete this task, following a pure GUI route—opening WeChat, finding the group chat, scrolling the page, and recording content. The content it summarized was relatively complete, including clear timestamps, and it could even clearly read the article revision suggestions in the images.

From the two cases, it can be seen that the combination of Hy4 + WorkBuddy has good reliability in office and long-horizon reasoning scenarios, and basically does not act on its own, meeting the rigor requirements of office scenarios.

3. Optimizing infrastructure to achieve a 20.1% throughput improvement, offensive training content already removed

Tencent also shared Hy4's capabilities in scientific research scenarios in the blog.

In AI research scenarios, Hy4 autonomously located performance bottlenecks in the pretraining system and carried out multiple rounds of infrastructure optimization across several directions, ultimately achieving a 20.1% improvement in end-to-end training throughput.

In the internal TenureTrail evaluation covering 7 major disciplines and 123 executable research tasks, Hy4 demonstrated solid scientific research capabilities in areas such as mathematical construction, physical modeling, Earth environment, systems engineering, and socioeconomic decision-making, ranking first among open-source models.

With Hyra 2.0, Hy4 developed the single-cell representation foundation model HyCell. On 7 cell atlases not involved in training, HyCell, with only 16 million parameters, achieved comprehensive performance comparable to the current SOTA model Stack with about 1/14 of the parameter count, and obtained the highest average biological fidelity score among the evaluated models.

In the direction of physical field modeling, Hy4, together with Hyra 2.0, developed the physics foundation model HyPDE. Compared with the current SOTA model Walrus, HyPDE achieved lower prediction errors on 9, 10, and 8 datasets out of 13 physical simulation datasets in single-step, short-term (1–20 steps), and long-term (21–60 steps) prediction, respectively.

Tencent has also open-sourced the above two models.

In terms of AI safety, Tencent said it has recognized the potential safety risks of the model's cyber and biological capabilities. During Hy4's training phase, the R&D team included only defensive cybersecurity data related to PoC generation and vulnerability remediation, while removing offensive training content such as vulnerability exploitation, reverse engineering, and penetration, striving to avoid improvements in potential offensive capabilities.

In biology, Hy4's open-ended experimental design capability still has a clear gap compared with frontier models. Tencent organized internal experts to audit trajectories and found that the model has strong mastery of basic biological knowledge, but its ability to translate knowledge into executable experimental plans remains limited, and it has not been found to significantly lower the threshold for carrying out biochemical harmful behavior.

Conclusion: Pretraining and post-training scale will continue to increase

In the blog, Tencent mentioned that Hy4 still has a considerable gap compared with top closed-source models, especially on the high-difficulty test sets built internally by Tencent in areas such as frontier mathematics, long-horizon development, model iteration, and visual question answering.

In the future, Tencent will continue to steadily expand the scale of pretraining and post-training, striving to narrow the gap with top closed-source models.