Tencent Hunyuan Hy4 and Alibaba Qwen4 to Face Off in Same Generation Next Week

By Letter List

Just as Tencent announced that the official version of Hy4 is about to be released, Alibaba put out word in Europe that Qwen4 will take the stage next week.

This is not the first PK between Tencent and Alibaba, but it will be the first head-on showdown between the two major models, Qwen and Hy.

In the past, the two companies' large models often adopted a staggered strategy of one advancing while the other retreated: you release a model first, and after a while I release mine. This time, however, the two even share the same version number, a strong signal of confrontation.

In September 2025, Alibaba launched the official version of Qwen3-Max, while Tencent's Hy2 was not released until December 5 of that year.

In 2026, the situation reversed. Tencent first released the official version of Hy3, rebuilt from Hunyuan, on July 6. Although Alibaba released Qwen3.8-Max on August 3, its major version number remained at 3.

Next week, all the conditions will be in place: same-generation official flagships, the same release window, and the same bet on Agents and real productivity.

For Liu Dayiheng, this is not an ordinary new product launch. Qwen4 is the first complete generational model since he formally took over, and also his first battle as the new leader of Qwen.

In May 2026, Ma Huateng once commented on Tencent's AI this way: "A year ago we thought we had gotten on the boat, then we found that the boat was leaking. Now it feels like we've stood up on it, but still can't sit down. We still hope the boat can go a bit faster."

In October, when Guangdong provincial leaders came to Tencent's headquarters for a research visit and discussion, Yao Shunyu was sitting right beside Ma Huateng. It seems Yao Shunyu has repaired this originally "leaking boat" enough for Ma Huateng to sit down in it.

Can Liu Dayiheng win?

01

It is still too early to draw conclusions now, but what is certain is that Liu Dayiheng cannot afford to lose.

In the past, competition between Tencent and Alibaba mostly took place in products, cloud, and ecosystems. From DingTalk and WeCom to Qianwen Office and WorkBuddy, the two companies were fighting over entry points: whoever could get into an enterprise's workflow would have the chance to turn AI into real productivity.

But on flagship foundation models, the two sides have always been one step apart. This frequency mismatch also gave both sides enough room for their own narratives.

Alibaba could talk about Qwen's model capabilities, open-source ecosystem, and global developers, while Tencent could talk about how Hunyuan connects to Yuanbao, CodeBuddy, and WorkBuddy, and how it uses its own product scenarios to feed back into the model.

It is not that the two sides have not competed; it is just that they have never faced off at the same time with same-generation official flagships.

The smell of gunpowder began to spread at the end of August.

On August 26, Alibaba released Qwen3.8-Flash-Next, previewing the Qwen4 architecture in advance. On August 28, Tencent released and open-sourced Hy4 preview.

The two releases were only two days apart, and both placed Agent and real productivity in important positions.

But this was only a probe between preview versions, and at that time Liu Dayiheng had not yet publicly appeared as Qwen's new head, so it could not yet count as a head-on showdown between him and Yao Shunyu.

It was equivalent to the two sides showing each other their opening stances before a martial arts contest, to see whose horse stance was more stable.

When it comes to Qwen4 and Hy4, both being main digital generations, Liu Dayiheng and Yao Shunyu likewise carry the pressure from their respective organizational adjustments. The two will be placed in the same coordinate system by the market for the first time.

Liu Dayiheng is someone who fought his way up inside Alibaba.

After joining DAMO Academy in 2021, he was an early core member of Qwen, long responsible for pre-training, and later his responsibilities expanded to post-training and Coding, and he also participated in the research and development of multimodal, mathematics, and other models.

Liu Dayiheng is a "new" person in front of Yao Shunyu, because this is the first time Liu Dayiheng has to be responsible for the overall result of this model.

In the past, doing pre-training well was enough to complete the KPI. From now on, he has to shoulder everything: architecture, post-training, Agent capabilities, and the open-source narrative.

After Lin Junyang left, what Alibaba needed was not just a model, but also to let the market know that Qwen's system can still operate stably to this day.

Yao Shunyu and Liu Dayiheng are completely opposite. He parachuted into Hunyuan with the most top-tier resume.

He is a co-proposer of ReAct and Tree of Thoughts, a former OpenAI researcher, and also participated in projects such as OpenAI's Operator and Deep Research.

After Yao Shunyu came to Tencent, he rebuilt the pre-training and reinforcement learning infrastructure, adjusted the R&D organization, and connected the model with product feedback. By July 2026, Hunyuan's multimodal model department and large language model department were merged into the Foundation Model Department, under Yao Shunyu's unified management.

Hy4 was born precisely against this background. This model is a product that fully implements Yao Shunyu's will.

The reason Liu Dayiheng's first battle is difficult is not only because Yao Shunyu is famous, but even more because Yao Shunyu has already transformed his personal research capability into organizational capability.

Of course, Yao Shunyu likewise needs the official version of Hy4 to answer: can vision and Agent truly be combined? How much performance improvement can full-link RSI actually bring? And can Tencent's huge product system continue to feed back into the model?

02

Having discussed the people, let us return to the models themselves. Neither Qwen4 nor Hy4 is a new product without groundwork; both sides have already revealed part of their cards in the preview versions.

First, Qwen4. Its biggest highlight is that the new architecture is being tested at the flagship level for the first time.

Alibaba explicitly defines Qwen3.8-Flash-Next as an early preview of the Qwen4 architecture.

This model combines Gated DeltaNet (GDN) with Qwen Sparse Attention (QSA) in its attention mechanism, while also introducing gated residuals, N-gram Embedding, and the Muon optimizer.

The first highlight is the level of intelligence.

According to Alibaba's disclosure, its backbone has 125B parameters, plus another 51B N-gram Embedding, with 6B parameters activated per Token, and its training cost is only one-ninth that of Qwen3.7-Plus.

Liu Da Yiheng cares deeply about "how much intelligence can each unit of compute actually buy?"

Previously, the new architecture first validated feasibility on Flash. Now, with Qwen4, the question it must answer is whether this architecture can carry over to Max.

Because Max typically has to handle larger-scale and more complex tasks, what works for lightweight models can easily fail when scaled to models with trillions of parameters.

If it can only produce cheap small models but cannot raise the intelligence ceiling of the flagship, that would show Liu Da Yiheng still falls a bit short.

The second highlight is real Agents.

Liu Da Yiheng's speech title at the Yunqi Conference was "Qwen: Toward Real-World Agents." The words "real-world" matter a great deal, because an Agent cannot merely call a few tools in succession in a demo; it must remember goals over long tasks, handle context, use tools, correct errors, and ultimately deliver directly usable code, documents, or analysis results.

If Qwen4's improvements show up only in single-turn Q&A and not in the success rate of long-horizon tasks, it will be hard to support Alibaba's narrative about real productivity.

The third highlight, and also the most distinctive line for Qwen4, is RSI.

At the Yunqi Conference, Liu Da Yiheng stated that Qwen3.8-Max can already, with almost "zero human involvement," autonomously build training pipelines, construct training data, design experiments, and locate defects, iterating continuously for more than a month and completing 33 effective iterations.

According to Alibaba's data, its Artificial Analysis evaluation score rose from 40 to 45.

Liu Da Yiheng said that RSI has also extended from model training to inference and chips. After the model autonomously adapted to Pingtouge's new architecture, the single-instance throughput of the inference framework increased by 96%. In chip co-design, the model, relying solely on the bus module specification, went through 60 hours of self-iteration and more than ten thousand EDA tool calls, reducing the physical implementation area by 42%.

He summarized this relationship as: "Models design chips, and stronger chips in turn feed stronger models."

Now let us look at Hy4. Tencent says that this time the focus is on strengthening long-horizon and real productivity task capabilities, significantly improving chain-of-thought efficiency, and introducing visual understanding for the first time.

The first two are the core of the Alibaba-Tencent showdown—essentially a contest of whose fist is bigger and whose arm is thicker.

Visual understanding is an important change in Hy4.

After all, real-world work is not only plain text; it also includes multimodal elements such as screenshots, charts, PDF layouts, web pages, and software interfaces.

With visual understanding introduced for the first time, whether Hy4 can move from "understanding the task" to "understanding the work site" and connect vision, tool calling, and long-horizon execution is an important difference between it and Hy3.

At the same time, Hunyuan has also, for the first time ever, launched Hy4 Flash, responsible for cost-effectiveness and large-scale invocation tasks.

Tencent says that Hy4 Flash will inherit Hy4's vision and Agent capabilities and evolve at high frequency through deep Co-Design between model and product.

This means Tencent not only wants a high-scoring flagship, but also wants to push capabilities down into higher-frequency real-world scenarios such as WorkBuddy and CodeBuddy.

Strictly speaking, though, this is not the first time Tencent Hunyuan has been tiered. As early as 2024, Tencent Cloud offered hunyuan-pro, hunyuan-standard, and hunyuan-lite, and later Turbo as well. HY2's Think and Instruct also corresponded to reasoning and instruction versions, respectively.

Hy4 Flash's first is that, after Yao Shunyu's rebuild, it is the first time the same generation clearly forms a dual-tier layout of "flagship + Flash."

Alibaba and Tencent have already entered the octagon, and inside the octagon, anything can happen.