Tencent releases 295B parameter Hy3 model to target AI agent efficiency

Tencent Holdings Ltd. on April 23 unveiled its Hy3 preview model, an open-source artificial intelligence system with 295 billion parameters, signaling a strategic pivot toward efficiency and practical application over chasing sheer scale. The launch represents a significant recalibration for the Chinese tech giant, focusing on creating a cost-effective model for complex, real-world tasks as the domestic AI race intensifies.

"Hy3 preview is the first step of the Hunyuan large model reconstruction," Yao Shunyuan, Tencent's chief AI scientist and head of its large language model division, said in a statement. "We hope that this open source and release will get real feedback from the open source community and users to help improve the practicality of the official version of Hy3."

The new model uses a Mixture-of-Experts (MoE) architecture with 295 billion total parameters but only activates 21 billion for any given task, along with a 256K context window. This design aims to balance high-end capabilities with lower operational costs. Internal tests show the model delivering a 54 percent reduction in first-token latency and a 47 percent decrease in end-to-end duration on internal products like the AI agents CodeBuddy and WorkBuddy, with a success rate greater than 99.99 percent.

The move suggests Tencent is betting that superior engineering and deep integration into its vast product ecosystem can provide a competitive edge, even without having the industry's largest model. This comes as Tencent and rival Alibaba Group are reportedly in talks to invest in DeepSeek, an AI startup seeking a valuation over $20 billion, indicating a dual strategy of building in-house while buying into promising third-party technology.

A Strategic Shift to Efficiency

The release of Hy3 marks a clear shift in Tencent's AI philosophy. The company now views the 300-billion-parameter range as an optimal balance point, where core capabilities like complex reasoning and long-context understanding are fully unlocked, and further increases in size yield diminishing returns. This counters the industry's earlier focus on ever-larger parameter counts as the primary measure of a model's power.

This recalibration follows a reorganization of Tencent's AI teams and the establishment of new infrastructure in February. The focus is now on "AI Agents" that can execute complex workflows, a concept Tencent's cloud and smart industry CEO, Tang Daosheng, has called the next paradigm. By open-sourcing Hy3 preview, Tencent aims to accelerate its evolution by gathering real-world usage data, refining the model for its formal release.

The Competitive Landscape

Tencent's pragmatic approach is set against a backdrop of intense competition and resource constraints. The high cost of training and scaling models is compounded by U.S. export restrictions on advanced semiconductors from companies like Nvidia, forcing Chinese firms to evaluate domestic alternatives from providers such as Huawei.

In this environment, efficiency is paramount. DeepSeek, the startup Tencent is reportedly courting, built its reputation on developing powerful models with a fraction of the budget used by Western counterparts. Tencent's pursuit of a stake in DeepSeek, while simultaneously launching its own efficiency-focused Hy3 model, highlights a two-pronged strategy: developing powerful, integrated in-house models for its ecosystem while using investments to gain exposure to other innovators and hedge against technological dead ends. The success of this dual approach will be critical in navigating China's uniquely challenging AI market.

This article is for informational purposes only and does not constitute investment advice.