AI News

Qwen4-Max Preview Released, Compared Alongside Tencent Hy4 in the Same Period

36Kr · AIUpdated Nov 13 21:44

AI summary

Qwen released the Qwen4-Max preview on November 11 and announced that Qwen4 will be officially released soon; Alibaba-affiliated products have already integrated it, while the official version has not yet launched. The author used the same prompt to have the Qwen4-Max preview and Tencent Hy4 each produce a version of a web page and a 3D animation, concluding that Hy4 is richer in visual completeness and interaction detail, while the Qwen4 version's visuals converge rather coarsely. The article also mentions that on the GameFeel Bench internal beta leaderboard, Qwen-4-Flash-Early scored 49.9 points, but the leaderboard has only 12 stars and 4 of its 7 metrics rely on manual scoring, so it should be viewed with caution.

Why read this

AI score 61

The article compares the Qwen4 preview with Tencent Hy4 within the same time window and presents measured differences in web pages and 3D animation, offering a view of the trade-offs in the two companies' new-generation model approaches.

Topics

Tags

Body · AI translation

On November 11, Qwen released the Qwen4-Max preview version and announced that Qwen4 will be officially released soon.

Since October, news about Alibaba's Qwen4 has been circulating. Previously, an anonymous model named Kalmia appeared on the large-model blind-testing platform Arena and was widely speculated to be Qwen4; internal test results regarding its game development capabilities also circulated on social media.

Officials have also been continuously teasing it. On November 6, Qwen posted a poster — "In the Northern Sea there is a fish, its name is _." After the Qwen4-Max preview was officially announced, Alibaba-affiliated products such as Qwen Office integrated it at the first opportunity; the official version has not yet been released.

Qwen said that Qwen4 has deeper knowledge understanding and more three-dimensional world perception, and developer internal test results show that it performs impressively in 3D modeling, web design, interaction and other areas.

But in the large-model field during the same period, Tencent has already taken the lead. On November 4, Tencent Hunyuan, led by Yao Shunyu, released Hy4, claiming it ranks among the strongest open-source models. Next, it is Liu Da Yiheng's turn to deliver his answer. Whether the Qwen4 in his hands can take Qwen a step further in the new round of competition is worth watching.

01 Does Qwen4 want to chase highs with large parameters and fill gaps with small models?

On November 9, a game development creativity evaluation leaderboard called GameFeel Bench went viral in the community: the Qwen-4-Flash-Early API, codenamed hawk, scored 49.9 in internal testing, beating Opus 5.5 xHigh's 47.2; Qwen-4-27B-Early scored 44.6, only 0.4 points behind Sonnet 5.5.

It is said that the testing team obtained internal test API keys before Qwen4's official release, covering the Flash version and the 27B deployment version. However, this third-party leaderboard still needs to be viewed with caution. GameFeel Bench has only 12 stars on GitHub, and 4 of its 7 evaluation metrics rely on manual scoring, so the results are inevitably affected by subjective judgment.

(Left/Qwen4-Max preview; Right/Hy4)

About 10 minutes later, the Qwen4-Max preview delivered a version of the homepage. The page structure was complete, and it also provided two themes, "Night Flight" and "Paper." However, the first screen did not reserve a section for the latest news, the headline in the hero area occupied nearly one-third of the screen, and there was no prompt to scroll down, so readers might leave after staying on the first screen. Overall, as a first draft, this homepage version was relatively complete.

We used the same prompt to have Hy4 create a version as well; the framework and color scheme changed considerably, and the card on the right changed from brand-related indices to an introduction to the company's methodology. For this particular output, the Hy4 version had a higher degree of visual completion, and interaction details such as mouse following, card color changes, and scroll-down browsing were also richer (dynamic screenshots such as scroll-down browsing are not shown).

This kind of basic web design is actually hard to use to distinguish the upper limits of the models, so we then tried to have the models visually express "identifying a signal in noise": starting from a cluster of disordered particles that gradually converge into a directional structure, generating a 3D animated webpage of about 5 seconds.

Made by Qwen4-Max preview

Made by Hy4

In terms of visual effect, the image generated by the Qwen4-Max preview converged rather roughly, and the colors at the end were overexposed; Hy4 expressed the process of identification through a rotating light effect moving from left to right, which was relatively more refined.

Leaderboards and hands-on tests can only reflect part of a model's capabilities, and the above comparison also has subjective aspects. To understand the technical changes behind these performances, one still has to look at the architecture. Alibaba open-sourced Qwen3.8-Flash-Next on August 26 as an architecture preview of Qwen4. Its technical report shows that its number of activated parameters is about one-third that of the previous-generation flagship Qwen3.5-397B-A17B, and both its training token count and training compute (FLOPs) are about one-ninth of that model's. It uses less compute to achieve test results close to or even exceeding those of the previous-generation flagship.

Large parameters and small models mean two different things to Qwen.

From Qwen3.8-Max's 2.4 trillion parameters, to the industry's guess of 3 trillion for Qwen4, and then to the 5 trillion to 10 trillion planned for Qwen4.5 and Qwen5, joining the top tier is inseparable from the parameter race. At the same time, small and medium-parameter models such as 1.5B, 3B, 8B, and 27B are still being continuously open-sourced, extending capabilities to places the closed-source system cannot cover. Liu Da Yiheng explained at the recent Wuzhen Summit: leading closed-source companies will not focus on niche areas such as minority languages, niche vertical applications, and adaptation to special hardware chips. These shortcomings need to be filled by open-source small and medium-parameter models.

02 Launched at the same time, what to watch?

Liu Da Yiheng, born in 1993, and Yao Shunyu, born in 1998, are both "post-90s" technical executives. One rose through Qwen's early pretraining all the way up and, after taking over, is for the first time responsible for an entirely new generation of models; the other parachuted from OpenAI into Tencent and directly oversees core R&D work on Tencent Hunyuan large models and AI infrastructure.

The two companies' all-new generation models also happen to clash in timing. As early as the end of August, on the 26th Alibaba released Qwen3.8-Flash-Next to reveal the Qwen4 architecture in advance, and two days later Tencent released and open-sourced Hy4 preview, so the showdown between the two companies' new models had already begun.

At the September Yunqi Conference, Liu Da Yiheng, who took over as head of Qwen large models, announced the technical direction of Qwen4: model RSI self-evolution, architecture updates, and multimodal interaction, and confirmed that Qwen4 was already in training.

Among these, the model self-improvement item means introducing recursive self-improvement (RSI) into Qwen4's training process. The longer the task and the more steps there are, the richer the real feedback data the model accumulates during execution, and the training system then has the opportunity to screen from it materials that can be used for the next round of training.

However, RSI is not unique to Qwen; Tencent Hunyuan is also exploring a similar path. In October, Tencent's Hy Frontier team, together with relevant scholars, proposed the Terminal-RSI training framework to continuously improve the execution capabilities of terminal agents through recursive training.

As models compared in the same period, in terms of long context, the technical choices of Hy4 and the Qwen4 preview architecture are quite representative. Hy4 adopts Gated DeepSeek Sparse Attention (Gated DSA) and reuses sparse indices across layers through IndexCache; the Qwen4 preview architecture mixes GDN and QSA layer by layer, with about three-quarters of the layers using recurrent state compression and the remaining layers responsible for sparse attention retrieval. The former reduces the overhead of repeatedly computing indices within the sparse attention mechanism; the latter combines recurrent state compression with sparse retrieval to reduce the number of layers that need to perform global attention.

Beyond technical routes, another battlefield the two companies compete on is who can first enter real work scenarios. The omni-modal capability emphasized by Liu Da Yiheng is not limited to "seeing images and understanding videos"; it also includes combining visual perception with Agent actions so that the model can continuously execute tasks based on information in the real environment.

Tencent Hunyuan's Hy4, meanwhile, introduces visual understanding capability for the first time and has Computer Use capability, which can combine command lines and graphical interfaces to simulate a real person operating a computer to complete office work, development, and other tasks.

Model architecture shows the model's approach to processing information, but for the soon-to-appear Qwen4, it needs to prove its capability in actual tasks. Returning to scenarios, users only look at whether the model can reliably call tools, complete complex long-horizon tasks, and deliver results that meet requirements.

This article comes from the WeChat public account "Fourth Wave," author: Chen Junyi, published by 36Kr with authorization.

Source

36Kr · AI