Tencent's Hunyuan Hy3 model with 295B parameters targets enterprise AI

(Bloomberg) -- Tencent Holdings Ltd. has released its latest foundational AI model, the Hunyuan Hy3 preview, a 295-billion parameter model aimed at enterprise applications, signaling a focused strategy on practical, cost-effective AI solutions in China's competitive market. The move comes as Tencent and rival Alibaba Group Holding Ltd. are reportedly in talks to invest in AI startup DeepSeek, highlighting the intensifying race for AI dominance.

"Hy3 preview is the first step in the reconstruction of the Hunyuan large model," said Yao Shunyue, a former OpenAI researcher who is now leading the project at Tencent. "We hope to get real feedback from the open source community and users."

The new model has 295 billion total parameters but only 21 billion are active, a design choice that points to a focus on efficiency and real-world business scenarios. The model's core strengths are in Agent and Coding capabilities, areas where Yao has deep expertise. On Tencent's cloud platform, the Hy3 preview is priced as low as 1.2 yuan ($0.17) per million tokens for input, a move designed to attract a wide range of developers and corporate clients.

The launch of Hy3 is a significant step for Tencent as it rebuilds its AI strategy. The company is emphasizing a "co-design" approach, working closely with its various product teams to refine the model's performance in real-world scenarios. This strategy contrasts with the more research-oriented approach of some competitors and aligns with the broader industry trend of seeking profitable applications for AI. Tencent's move also reflects the growing importance of open-source models in the AI landscape, with companies like DeepSeek gaining traction by offering low-cost, open-source alternatives to proprietary models from giants like OpenAI and Anthropic.

The competitive landscape in Chinese AI is heating up, with Tencent and Alibaba not only developing their own models but also investing heavily in startups like MiniMax and now potentially DeepSeek. This frenetic pace of development and investment is happening despite China's limited access to the most advanced semiconductor technology, a challenge that makes the performance of models like Hy3 and DeepSeek's even more notable. As the "agent era" of AI dawns, the ability to provide not just powerful models but also the efficient hardware to run them, like Google's new TPU 8t and 8i chips, will be crucial. Tencent's focus on a smaller, more efficient model with Hy3 could be a savvy move in a market where cost and practicality are becoming increasingly important.

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