Tencent’s Dual AI Strategies: Hunyuan’s Scalable Ambitions vs. WeChat WeLM’s Efficiency Focus

Tencent’s AI journey is not a single track but a tale of two distinct initiatives, each with its own leadership, goals, and technical approaches.

On one side is the Hunyuan large model, led by Yao Shunyu, a 29-year-old (born in 1997) leader who joined Tencent half a year ago and brought in top industry talent. The latest Hunyuan Hy3 official version has been integrated into internal products like WorkBuddy and Yuanbao, narrowing the gap with leading models. Yao’s team has shifted away from chasing benchmark rankings or ultra-large parameters; instead, they focus on medium-scale parameters combined with optimized inference engineering to steadily enhance model capabilities. Hunyuan falls under the Technology Engineering Group (TEG), with Yao reporting directly to Tencent President Martin Lau.

On the other side is WeChat’s AI transformation, driven by long-tenured team members. The WeChat AI team, part of the WeChat Business Group (WXG), has been developing its own WeLM large language model since 2022—even before ChatGPT’s public launch in November 2022. This year, the team has released a series of technical blogs and is supporting the internal testing of 'Xiaowei,' WeChat’s first native AI assistant set to launch in Q3 2026. The team’s mission, as stated in their blog, is to 'explore the boundaries of intelligence under extreme resource efficiency.'

Internal resource competition is inevitable. According to Caixin, Tencent’s GPU resources are concentrated on Hunyuan and the Yuanbao app, leading to shortages that slowed WeLM’s training. The WeChat team had to repeatedly apply for more GPU cards from the group, and it took months to complete training.

WeLM’s approach to efficiency is rooted in practical needs. As a national app with over 1.4 billion monthly active users, WeChat faces enormous computing power cost constraints—even a tiny per-token cost difference multiplied by billions of users becomes a huge expense. To address this, WeLM has adopted innovative techniques: their July 2024 technical report 'Hidden Decoding at Scale' splits each token into four hidden computation streams, allowing capability improvements without major changes to the backbone network. This reduces training cost increases significantly—for example, an 80B model’s cost only rises 5.1x, far less than the 16x theoretical cost of naive scaling. Another blog on sparse MoE models mentions an 80B total parameter model with only 3B active parameters per step, further optimizing resource use.

The strategic question for Tencent’s management remains: will Hunyuan become the unified AI base across the group, or will the two parallel paths (scale vs. scene efficiency) continue? Both teams are building core base models, with external AI talent leading Hunyuan and in-house veterans driving WeChat’s AI efforts.

Sources

• OFweek AI Network
• Caixin
• LatePost