Tencent is pursuing not just one, but multiple AI strategies.

Written by | Zhao Weiwei

Tencent: The ‘Crouching Tiger,’ WeChat: The ‘Hidden Dragon’

On one front, Tencent Group’s top-level strategy is evident: Yao Shunyu, a ‘newcomer’ born in 1997, parachuted into Tencent, recruited top talent from the industry, and has been refining the Hunyuan large model for the past six months. The latest official version, Hunyuan Hy3, has been integrated into Tencent’s internal products such as WorkBuddy and Yuanbao, narrowing the gap with leading models. Their aim is to demonstrate that while Tencent may have entered the race late, it still has a strategic rhythm.

On the other front, AI transformation is underway within the WeChat ecosystem: Led by a Tencent veteran with over 20 years of experience, Zhou Jie and a team of dedicated WeChat AI members have been developing the self-researched WeLM large language model since 2022. This year, they have frequently updated their technical blogs to support WeChat’s internally tested AI assistant, ‘Xiaowei,’ which is planned for launch in the third quarter of this year. Their goal is to show that a self-developed, cost-effective, and efficiency-first model is better suited to WeChat’s ecosystem scenarios.

‘No wonder it’s Tencent; even large language models (LLMs) have two competing teams,’ commented a user on the social media page of WeLM’s technical team.

‘We are committed to pushing the boundaries of intelligence under extreme resource efficiency,’ wrote the WeChat AI team in their blog. This sentiment also mirrors Yao Shunyu’s approach after joining the Hunyuan team. He urged Hunyuan to abandon the pursuit of benchmarks, steering clear of trillion-parameter scales or ultra-long contexts, and instead focused on steadily enhancing model capabilities through medium-scale parameters and optimized inference engineering.

The key difference lies in stability: After three generations of technical leaders, the Hunyuan large model team has finally stabilized, transitioning from Zhang Zhengyou to Jiang Jie, and now to Yao Shunyu. Hunyuan falls under the Technology Engineering Group (TEG), with Yao reporting directly to Martin Lau.

Meanwhile, WeChat AI’s WeLM belongs to the WeChat Business Group (WXG), led by a stable, internally grown team headed by a founding member of the WeChat team.

WeChat’s AI transformation is spearheaded by local ‘veterans’ who intimately understand the product, narrating the story of how a product with 1.4 billion users remains low-key, restrained, and deeply integrated into AI scenarios.

Hunyuan’s transformation, on the other hand, is led by the high-profile ‘newcomer’ Yao Shunyu, with a narrative centered around building a powerful base model to drive growth in Tencent Cloud and gaming businesses.

Tencent holds more than one AI ‘ticket.’ Hunyuan and WeLM represent two distinct routes: one pursuing scale breakthroughs, the other delving into in-depth scenario development. Both paths point to Tencent’s future of sustained investment in AI.

Both the WeChat AI and Hunyuan teams are constructing core base models. The question remains whether external AI newcomers and WeChat’s local veterans will operate in parallel or strategically converge. Will Hunyuan emerge as Tencent’s unified AI base?

This is a pressing issue for Tencent’s management.

1 WeChat AI: A Low-Key Veteran Team

In contrast to the significant attention Yao Shunyu receives for Tencent’s Hunyuan large model, the leader of WeChat’s AI team remains notably low-key, rarely making public appearances in recent years.

He currently serves as the technical architecture leader of the WeChat Business Group, overseeing the internal testing of WeChat’s intelligent agent, ‘Xiaowei,’ WeChat’s first native AI assistant, slated for a wider user release in the third quarter of this year.

His most widely recognized public appearance was a decade ago in a presentation titled ‘The Architectural Secrets Behind 100 Million User Growth,’ where he proposed four architectural principles—‘build large systems small, make everything extensible, must have foundational components, easy to launch’—principles that have since been repeatedly cited in discussions about WeChat’s high-concurrency systems.

Less known is that WeChat WeLM’s first technical paper, ‘WeLM: A Fully Pre-Trained Chinese Language Model,’ published in September 2022, predates ChatGPT’s official release in November 2022, making it one of the earliest domestic teams to follow the general pre-trained large model route.

At the time, WeChat AI trained the model with 10 billion parameters, significantly outperforming pre-trained models of the same scale and rivaling models 25 times larger.

The paper listed eight authors, including Su Hui, Zhou Xiao, Hou Jinyu, Shen Xiaoyu, Chen Yuwen, Zhu Zilin, Yang Yu, and Zhou Jie (phonetic translations).

Currently, except for two individuals confirmed to have left via public records, WeChat AI’s original large model team has remained largely stable over the past four years.

Su Hui, one of the first authors, joined Meituan in 2023 to develop Meituan’s LongCat-Flash; Zhu Zilin joined Zhipu in 2025 as an RL Infra engineer, leading the open-source framework ‘slime’ for post-training reinforcement learning in LLMs.

The July 2026 paper reveals that WeLM’s base large model team has grown to 48 members, a sixfold increase, with Chinese names dominating and only one author, Donald He (Donald · He), suspected of having an overseas background.

The last author in model papers typically indicates the leader of the research direction.

Zhou Jie, the most frequently mentioned in WeLM’s technical papers, is the head of WeChat’s Pattern Recognition Center and WeLM’s technical leader. He guided the team through full-series technical iterations from 10B dense models to MoE architectures, serving as the top technical supervisor.

Zhou joined Tencent in 2017, previously a Ph.D. from the Institute of Theoretical Physics, Chinese Academy of Sciences, later shifting to computer deep learning and AI, with a focus on natural language understanding. He remains very low-key, last publicly appearing at Hunan University’s general education lecture in June 2024.

Clearly, from WeChat AI’s leader to WeLM’s technical leader and the stable large model team since 2022, WeChat AI is dominated by long-time Tencent ‘veterans.’

Unlike Yao Shunyu’s Hunyuan large model, which brought in many ‘newcomers,’ WeChat AI follows an internally grown, stable iteration route, with deeper team roots, lower turnover, and tighter alignment with WeChat’s business scenarios.

2 ‘Frugality’ as a Core Principle

‘We are committed to exploring the boundaries of intelligence under extreme resource efficiency’ means that while others stack parameters with computational power, WeChat AI saves computational power with ingenuity.

WeChat AI must navigate internal competition. Caixin reported that Tencent’s computational resources are concentrated on the large model Hunyuan and Yuanbao App, with GPU shortages constraining WeLM’s training speed. The WeChat Business Group had to repeatedly apply to the group for new card resources, ‘taking months to run.’

Meanwhile, Tencent President Martin Lau believes that developing intelligent agents in WeChat first requires ensuring user privacy and security. Some scenario-specific functions cannot be easily achieved with general models. For WeChat’s massive user base, the model needs further inference capabilities.

Additionally, a Tencent insider interviewed by LatePost said Hunyuan now aims to break through intelligence limits, while WeChat seeks more economical computational modes to suit large user volumes.

WeChat AI’s choice of ‘frugality’ is essentially a commercial necessity.

WeChat, a national-level app with over 1.4 billion MAUs, faces core constraints from large model computational costs. Even a minuscule difference of 0.0001 yuan/1,000 tokens, when multiplied by a billion users, becomes enormous. This is why WeChat prioritizes cost optimization and insists on self-developed base models.

Doubao provides a cautionary tale, with rising computational costs forcing it to charge. According to a Guolian Minsheng Securities report in May, even the cheapest Doubao model costs between 132 million to 240 million yuan per day for free AI services.

How does WeChat AI address this? The July 14 technical report ‘Hidden Decoding at Scale’ offers insights.

When a base model’s training converges, how can its capabilities be further enhanced? The industry’s standard answer is path-dependent: expand parameters, add layers, widen models, and rerun full pre-training with ten-thousand-card clusters—linear computational investment for sub-linear performance gains.

WeLM’s approach is to split each token into four ‘hidden computation streams,’ allocating more internal computation per token while keeping the backbone network largely unchanged, thus continuing to improve model capabilities. As a result, training costs for the 80B model rose only 5.1 times, and for the 617B model, only 4.4 times, far below the theoretical 16 times cost of naive implementations.

Another technical blog, ‘Building Efficient Sparse MoE Models with Moderate Resources,’ describes an 80B total parameter, 3B single-step activated MoE model trained with under 14T tokens, matching or surpassing general models of similar or larger scales in performance.

The conclusion is straightforward: ‘These methods provide a practical roadmap for developing efficient and powerful models without relying on extreme-scale resources.’

This has been WeLM’s consistent approach over the past four years: in 2022, matching a 25 times larger opponent with 10B parameters; in 2026, supporting 1.4 billion users with 3B activated parameters. The technical blogs list numerous engineering optimizations that, combined, form a complete ‘extreme efficiency’ system—the art of ‘frugality.’

The scale of the problem has changed, but the underlying logic remains: solving scheduling for billion-user products with extreme resource efficiency.

From current tests of WeChat’s AI assistant ‘Xiaowei,’ Moments and private chat data can only access the last two days’ records, avoiding high computational costs for full historical data. Mini-program controls rely on official Skill APIs rather than visual recognition simulations. The Contacts Skill explicitly prohibits multiple tool calls in one round or parallel tool calls.

These reflect WeChat AI’s methodology for full-link cost control: not pursuing extreme usability but ensuring stable availability.

3 What Lies Ahead After Taking Center Stage?

Tencent’s current AI strategy has completed a four-tier layout:

The base Hunyuan large model provides reasoning, coding, and agent capabilities, aiming for the industry’s top tier; Yuanbao, as a general application, validates C-end demand and content consumption capabilities; WorkBuddy and CodeBuddy create closed loops for office scenarios and enterprise consumption; WeChat AI’s Xiaowei serves as a platform AI gateway, closing loops for WeChat content, advertising, and commercialization capabilities, raising the ceiling for AI narratives.

Hunyuan Hy3 has been released, and WeChat AI’s Xiaowei is in internal testing. Both newcomers and veterans are taking center stage.

The question is whether Hunyuan and WeLM, both building base large models, will cause internal product coordination issues. Goldman Sachs analysts directly stated in a report, ‘This means Tencent is maintaining two independent large model R&D systems in parallel. Whether they can achieve synergy and whether training costs can ultimately be integrated remains uncertain. The market worries this is redundant resource investment.’

Goldman Sachs estimates this investment could equal 5-17% of Tencent’s adjusted operating profit in the fourth quarter of 2026. Even so, they maintain a ‘buy’ rating, believing long-term ecological dividends will cover short-term redundant construction costs.

This argument about redundant resource investment views the issue through past experiences. Google’s 2023 merger of DeepMind and Google Brain provides valuable industry insight for observing Tencent’s dual-model system, validating a pattern: when large models become a company’s core strategy, redundant bottom-layer technology construction is eventually consolidated by scale effects.

But for Tencent now, it’s not an either-or choice. Dual-track parallel operations, while inefficient, preserve business-side flexibility. The only catalyst that might force a true merger is cost—profitability during the investment phase.

After all, the AI race has entered the ‘capital finals.’

Across the ocean, Anthropic’s valuation has surpassed OpenAI’s, reaching $965 billion, with a potential Nasdaq listing as early as October; OpenAI’s IPO plans have been delayed from this fall to 2027. On this side of the Pacific, Tencent-invested DeepSeek is also preparing for an IPO, valued at around $71 billion, with a potential 2027 listing.

Large models are a financially draining race where those who slow down in R&D investment may fall behind in the next model competition.

Reviewed by | Chen Qiulin

END