Yao Shunyu has barely settled into his seat.

Questioning him now is premature. He took over HunYuan in December 2025, and his first public address came in early June — six and a half months at most. For a massive system that needs rebuilding from three dimensions — organizational structure, data pipeline, and compute scheduling — half a year is far too short.

The restructuring has just begun; the weaknesses lie in the supporting infrastructure

Hy3 Preview is merely the first half-baked deliverable Yao has submitted, far from a finalized main model. He himself admits: “We are still scaling up pre-training and post-training. The official release will see major iterations.”

What he inherited was not a crack team but a pile of accumulated historical baggage. Tang Daosheng conceded publicly that in the past, the HunYuan team was scattered across multiple business lines with no unified foundation; high-quality training data supply was severely insufficient; and training GPUs were long prioritized for internal products, leaving compute resources tight.

What Yao Shunyu has accomplished in these six months is consolidating the fragmented workforce, building new data cleaning and evaluation pipelines, and pushing for compute expansion and scheduling. All of these efforts take months, and even now, the data infrastructure is not fully operational, while compute resources still cannot match the abundant resources of ByteDance and Alibaba.

So judging the entire system based on an interim version is unreasonable. Industry analysts point out that Alibaba’s Qwen 3.7 and ByteDance’s Doubao 2.1 both required over two years of iteration to stabilize. Expecting a team that has only been integrated for half a year to deliver instant breakthroughs — the time gap is simply too obvious.

Does Tencent actually know how to train models? It certainly does.

Yao Shunyu has demonstrated the full-chain technology on the spot. Hy3 Preview adopts the mainstream MoE architecture, fully aligned with the approach of ByteDance, Alibaba, and DeepSeek. He is the original proposer of the ReAct architecture and was deeply involved in OpenAI’s GPT training — there is no blind spot in his methodology.

The real bottleneck is execution resources

It wasn’t until late 2025 that Tencent began mass-recruiting experts in hundred-billion-parameter NLP pre-training and MoE, and the full talent pipeline is still being built. WeChat and QQ’s social data are tightly regulated and cannot be fully opened, making the scale and structural quality of high-quality corpora far inferior to Alibaba’s e-commerce data and ByteDance’s public-domain assets.

Tang Daosheng has admitted multiple times that resources are prioritized for products, leaving the large model’s dedicated compute resources tight. Both single-card utilization and cluster scale struggle to support daily iteration.

Battlefield intensity at maximum: iterations measured in days, competitors already on the assembly line

But the battlefield waits for no one. By 2026, the iteration pace has been compressed to its limit.

Industry iteration cycles have shrunk to 15-30 days — four to six times faster than 2023. And it runs year-round: new versions ship through Spring Festival and Dragon Boat Festival alike, with R&D working 7x24 in shifts. IDC data is brutally clear: miss one iteration cycle, and rankings and API call volume drop immediately. “Daily fine-tuning, monthly releases — the intensity is unprecedented.”

Tencent’s competitors are already truly in their element. ByteDance’s Doubao 2.1 launch showcased a complete pipeline: pre-training expanded with multimodal data, post-training automated clusters completing custom fine-tuning in 18 hours, weekly rolling updates, and a 315-million MAU product feeding real user feedback back into the model. The closed loop is already in place.

Tencent, meanwhile, is still transitioning from semi-manual to industrialized processes.

Time is running out for Yao Shunyu.

But to pass judgment on him now is still reading a eulogy before he has even warmed the seat. The systemic restructuring has only just taken shape; his capabilities have not been disproven.

In the coming year, he must move with maximum urgency to untie the ropes binding his hands and feet — fill the talent gaps, operationalize the data infrastructure, plug the compute shortfall, and turn “knowing how” into “doing it fast.”

AI is indeed a long-distance race. But the prerequisite is that you don’t first get dropped from the lead pack.

It’s time to pick up the pace.