Tencent's Hunyuan has launched its first major language model following a restructuring of the division. On April 23, the company released and open-sourced the Hunyuan Hy3 Preview language model. This hybrid expert model integrates fast and slow thinking approaches, featuring 295 billion total parameters, 21 billion activated parameters, and support for context lengths of up to 256K tokens. It is the first model trained after Hunyuan's reorganization and is described as the most intelligent model in the series to date, showing significant improvements in complex reasoning, instruction following, in-context learning, coding, agent capabilities, and inference performance.
This marks the first public disclosure of a major model from Tencent Hunyuan since its restructuring last year. In December, it was reported that Tencent had upgraded its large model development framework by establishing new departments for AI Infrastructure, AI Data, and Data Computing Platforms to strengthen its R&D system and core capabilities. Vinces Yao was appointed as Chief AI Scientist in the CEO/President's Office, reporting to Tencent President Martin Lau, while also serving as head of the AI Infrastructure and Large Language Model departments, reporting to Dowson Tong, President of the Technology and Engineering Group.
The release of Hy3 Preview represents the first major achievement for the 28-year-old AI prodigy since he assumed the role of Tencent's Chief AI Scientist. In September last year, Tencent officially refuted rumors circulated online claiming that former OpenAI researcher Vinces Yao had joined the company with a compensation package exceeding 100 million yuan.
According to public information, Yao is a graduate of the Yao Class at Tsinghua University, an elite program founded in 2005 by renowned computer scientist Andrew Yao to cultivate top-tier talent in computer science and interdisciplinary innovation. Yao previously served as co-chair of the Yao Class committee and was a co-founder of Tsinghua University's student rap club. In 2024, he joined OpenAI, where he contributed to the development of the agent product Operator and Deep Research, playing a key role in advancing language agent technologies. In May 2025, at age 27, he was named to MIT Technology Review's "Innovators Under 35" list for China, becoming the youngest honoree.
Tencent explained that in February, the Hunyuan team rebuilt the infrastructure for pre-training and reinforcement learning, establishing three principles focused on practicality: 1. Systematic Capability: The model avoids specialization in narrow tasks, as even single applications like code agents require deep coordination of multiple abilities including reasoning, long-text processing, instruction following, dialogue, coding, and tool usage. 2. Authentic Evaluation: Moving beyond leaderboards susceptible to manipulation, the team uses self-developed tests, recent examinations, human evaluation, and product crowd-testing to assess and improve the model's real-world performance. 3. Cost-Effectiveness: Practicality requires commercial viability. By deeply integrating model architecture with inference framework design, the team has significantly reduced task costs, making advanced AI more accessible and usable.
Public data indicates that coding and agent capabilities are the areas where Hy3 Preview shows the most notable improvements. Thanks to the rebuilt pre-training and reinforcement learning frameworks and increased scale of reinforcement learning tasks, the Hunyuan team rapidly achieved competitive results on major code agent benchmarks such as SWE-Bench Verified and Terminal-Bench 2.0, as well as search agent benchmarks including BrowseComp and WideSearch.
Hy3 Preview is already integrated with numerous Tencent products, with initial deployment on Tencent Cloud, Yuanbao, IMA, CodeBuddy, WorkBuddy, QQ, QQ Browser, Tencent Docs, and Tencent LeXiang. Services including WeChat Official Accounts, Peacekeeper Elite, Tencent News, Tencent Stock, Tencent Customer Service, and WeRead are progressively adding support. Additionally, Hy3 Preview supports integration with popular open-source agent products like OpenClaw, OpenCode, and KiloCode, and is available on Tencent Cloud's large model service platform, TokenHub.
Multiple evaluations show that the Hy3 Preview model improves overall inference efficiency by 40% while significantly reducing costs compared to the previous generation. On TokenHub, pricing for Hy3 Preview starts at ¥1.2 per million tokens for input, ¥0.4 per million tokens for cached input, and ¥4 per million tokens for output. Tencent Cloud has also launched a customized Hy3 Preview Token Plan, with a personal edition starting at ¥28 per month, offering cost-effective options for agent development and creating "Lobster" applications.
The release of Hy3 Preview has generated significant anticipation and intensified competition in China's AI sector. In April, besides Hunyuan's model, another highly anticipated domestic model is DeepSeek V4. Online rumors suggest V4 will launch in April, though DeepSeek has maintained secrecy about its specific progress despite recent updates.
On April 22, foreign media reported that Chinese tech giants Tencent Holdings and Alibaba Group were in talks to invest in DeepSeek, an AI startup that recently began its first funding round. An informed source indicated DeepSeek is seeking funding at a valuation exceeding $20 billion. However, another source explicitly denied these reports, calling them inaccurate. As of publication, Tencent and Alibaba had not commented on the matter.
Tian Feng, Dean of the Fast and Slow Thinking Research Institute, noted that domestic large models still have considerable potential for commercial valuation. According to Stanford University's "2026 AI Index Report," the performance gap between top Chinese and U.S. models is now only 2.7%. However, while OpenAI's annualized revenue exceeds $24 billion and Anthropic's reaches $30 billion, DeepSeek—focused on fundamental research—has not yet set commercialization targets or launched commercial products, resulting in a significant valuation disparity from a business perspective.
Comments