Tencent open-sources 770B-parameter Hy4 Preview, edges GLM 5.3

Tencent's Hunyuan division released Hy4 Preview, a 770B-parameter open-source model with 49B activated parameters and 1M-token context, built for productivity tasks spanning coding, office analysis, and scientific research.

"Hy4 Preview has significantly expanded in size, context, and data scale, and the simultaneous advancement of pre-training and post-training has led to another leap in intelligence level," Tencent said, noting the model now ranks among the top open-source models.

In blind tests organized by Tencent with 163 internal experts across 203 engineering tasks, Hy4 Preview scored 2.99 out of 4.00, slightly ahead of Z.AI's GLM 5.3 at 2.92 and Moonshot AI's Kimi K3 at 2.94. The model also advanced the three-dimensional Blaschke–Lebesgue problem, pushing the volume lower bound from 0.380799 to 0.41104 — within about 2 percent of the Meissner tetrahedron conjecture at 0.41986.

Tencent shares rose 2.1 percent to HKD457.2 on the announcement, while Z.AI fell 7.3 percent. Hy4 Preview is priced at RMB6 per million input tokens and RMB18 per million output tokens, with cached-hit tokens at RMB0.3 per million — a structure that could pressure rivals in China's LLM market.

Open-source distribution and product integration

Hy4 Preview is available on HuggingFace, GitHub, ModelScope, and Gitcode, and accessible through APIs via Tencent Cloud TokenHub and OpenRouter. The model is deeply integrated with CodeBuddy and WorkBuddy, Tencent's developer and workplace productivity tools, and is also available through Yuanbao and ima. WorkBuddy and CodeBuddy will run a two-week free access campaign to gather user feedback and refine performance.

The model's domain-specific strengths are notable. On the software engineering side, it strengthens long-term development understanding, planning, and verification. In office analysis, it can complete full delivery from data processing to document table presentation. For game development, it can generate playable prototypes from a single requirement. In research, it has made progress in molecular dynamics and condensed matter physics.

Competitive pressure in China's LLM market

The open-source approach and aggressive pricing put direct pressure on Z.AI and Moonshot AI, whose flagship models now face a well-funded competitor with deep integration into Tencent's product suite. Hy4 Preview was co-built using high-quality data from internal experts in software engineering, gaming, finance, and security, giving it domain-specific advantages in real-world productivity tasks.

Tencent acknowledged Hy4 Preview is an early iteration with known issues in complex task reasoning and excessive self-verification, suggesting the competitive environment remains fluid. The company said it will continue to iterate agilely on both pre-training and post-training.

The pricing strategy is particularly aggressive. At RMB6 per million input tokens, Hy4 Preview undercuts many comparable models in the Chinese market, and the RMB0.3 per million cached-hit rate makes repeated inference workloads significantly cheaper. This could accelerate adoption among developers and enterprises building on open-source models, potentially reshaping pricing dynamics across China's LLM sector.

Tencent's push into open-source AI strengthens its cloud services and productivity software revenue streams. The stock, trading at HKD457.2 with turnover of HKD8.046 billion, reflects investor optimism about the model's commercial potential. However, the competitive dynamics in China's LLM market remain volatile, and the two-week free access period will be a key test of user adoption and retention.

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