Tencent Hunyuan Hy4 Preview: 770B Open-Source Model Claims Top Tier

Model Specifications

On August 28, Tencent Hunyuan released and open-sourced Hy4 preview, a new generation large language model. The model uses a mixture-of-experts architecture with 770B total parameters and 49B activated parameters , and supports a context length of 1M tokens . Significant scaling was applied to model size, context length, and training data volume, with joint improvements in pre-training and post-training driving a major leap in intelligence, placing it firmly in the first tier of open-source models.

Benchmark Results: Internal Blind Test

Tencent conducted a blind evaluation with 163 internal experts across 203 engineering tasks . Hy4 preview achieved an average score of 2.99/4 , slightly outperforming GLM 5.3 (2.92/4) and Kimi K3 (2.94/4). The model is positioned as "born for productivity" and was co-developed with high-quality data from Tencent experts in software engineering, gaming, finance, and security, plus deep co-design with products like WorkBuddy.

Capability Domains

Software Engineering: Enhanced long-horizon development task understanding, planning, debugging, and verification; improved front-end visual aesthetics and interaction quality.

Office Analysis: Significantly improved complex office environment comprehension and financial analysis; optimized data analysis and cross-file collaboration, enabling end-to-end delivery from information processing to documents, spreadsheets, and presentations.

Game Development: Strengthened single-prompt generation of playable prototypes; proficient use of game engines; supports multi-turn iterative refinement of complex game projects.

Scientific Research: Notable gains in understanding, reasoning, and solving complex research problems; progress across AI R&D, molecular dynamics simulation, condensed matter physics, and fundamental mathematics.

Recursive Self-Improvement Loop

Hy4 preview participated in its own full-chain R&D process for the first time, contributing to training methods, data strategies, evaluation systems, and low-level operator auto-optimization. The model proposed solutions, ran experiments, and iterated based on results, with generated code, logs, and feedback feeding into the next exploration round — forming an initial recursive self-improvement closed loop .

Inference Infrastructure Optimization

The model autonomously analyzed inference system bottlenecks and conducted multi-round optimizations around operator fusion and communication improvements. This yielded a 31.8% end-to-end throughput increase over the baseline , with stable gains across varying context lengths and concurrency levels, demonstrating preliminary ability to self-optimize inference infrastructure.

Availability and Pricing

Hy4 preview is open-sourced and simultaneously launched in WorkBuddy/CodeBuddy (domestic and international versions), Yuanbao, and ima. API access is available via Tencent Cloud Tokenhub and OpenRouter. Pricing follows a cost-effective inclusive route: input 6 CNY/million tokens, output 18 CNY/million tokens, cache hit 0.3 CNY/million tokens . A 2-week free trial is offered on WorkBuddy/CodeBuddy to collect user feedback for further model improvements.

Release Cadence

Since rebuilding infrastructure in February, Hunyuan has averaged a major version iteration every two months. Using a preview-first, formal-release-follow-up approach, real-world feedback is continuously fed back into R&D. The next version of Hy4 is expected to roll out soon following this rhythm.