Technology & AIAnalysis
Tencent Open-Sources Hy4 Preview Model with 770 Billion Parameters
The company's latest sparse-activation model targets engineering tasks and features a context window exceeding one million tokens.
Share
Photo by mingche lee on Pexels
The Brief
Tencent has open-sourced Hy4 preview, its latest flagship large language model designed specifically for enterprise and technical productivity. Featuring 770 billion total parameters with 49 billion activated per token, the model boasts a context window exceeding one million tokens. According to Tencent, the model scored 2.99 out of 4.00 in blind evaluations across 203 engineering tasks conducted by 163 domain experts, marginally edging out rivals like GLM 5.3 and Kimi K3. The release marks Tencent's ongoing push to integrate foundation models directly into developer tools and corporate workflows.
Why it matters
Hy4 preview underscores Tencent's commitment to open-source foundation models that balance massive capacity with computational efficiency via sparse activation. By targeting code generation, scientific research, and enterprise office workflows, Tencent is positioning itself directly against leading Chinese frontier models while expanding its developer footprint through domestic and global application programming interfaces.
China context
China's generative AI landscape is shifting rapidly from raw benchmark competition to measurable productivity gains across software engineering, gaming, and enterprise operations. Tencent's approach leverages its vast internal business ecosystem—ranging from financial services to game production—to supply domain-specific training data and refine models before releasing them publicly. This pattern reflects a broader domestic trend of big tech firms using proprietary corporate infrastructure to fuel recurring open-source releases.
Editor's View
EDITOR'S VIEW — Analysis and inference, not factual reporting.
Tencent's decision to deploy a 770-billion-parameter mixture-of-experts architecture with over one million tokens of context illustrates how table stakes in Chinese open-source AI have expanded. Rather than keeping frontier capabilities proprietary, Tencent is pursuing a dual strategy: embedding Hy4 into internal products such as Yuanbao and WorkBuddy while distributing it via external platforms like OpenRouter and Tencent Cloud. While the self-reported blind test scores indicate competitive parity with domestic peers, the true measure of Hy4 will depend on independent community validation and operational inference costs at scale.
What to watch
- Independent evaluations and benchmark reproductions from the broader open-source developer community
- Adoption rates of Hy4 preview APIs on platforms such as OpenRouter and Tencent Cloud Tokenhub
- Whether Tencent sustains its targeted cadence of delivering major model iterations every two months
Key Takeaways
- 1Tencent released Hy4 preview with 770 billion total parameters and 49 billion active parameters.
- 2The model supports an expanded context window exceeding one million tokens.
- 3In company-reported blind testing on 203 engineering tasks, Hy4 preview scored 2.99 out of 4.00, compared to 2.92 for GLM 5.3 and 2.94 for Kimi K3.
- 4Hy4 preview was integrated into Tencent products such as WorkBuddy and Yuanbao, and made accessible via OpenRouter and Tencent Cloud.
- 5Tencent reported that the model participated in its own development pipeline through recursive self-improvement.
Tencent has released and open-sourced Hy4 preview, its next-generation large language model designed to handle complex workplace, coding, and scientific applications, according to state media outlet People's Daily. The launch represents a significant expansion in scale, context capacity, and architectural refinement for Tencent's Hunyuan model series.
The model features 770 billion total parameters, with 49 billion parameters activated dynamically during inference, utilizing a sparse architecture to manage compute requirements. Additionally, Hy4 preview expands its context window beyond one million tokens, enabling the processing of extensive codebases and lengthy technical documentation in a single prompt window.
In human evaluations reported by the company, Hy4 preview achieved an average score of 2.99 out of 4.00 in blind tests across 203 engineering tasks evaluated by 163 domain specialists. This result placed the model slightly ahead of comparable domestic frontier models, including Zhipu AI's GLM 5.3, which scored 2.92, and Moonshot AI's Kimi K3, which scored 2.94. These benchmarks were focused predominantly on technical scenarios spanning software engineering, financial analysis, gaming development, and security protocols.
Tencent noted that Hy4 preview was trained through collaborative data-building initiatives with specialists in specialized industry verticals. Notably, Tencent stated that the model was deployed in a recursive self-improvement loop during its own development, directly assisting in optimizing training techniques, data curation strategies, and internal evaluation frameworks.
The model has been rolled out across Tencent's proprietary product suite, including domestic and international versions of the developer-oriented WorkBuddy and CodeBuddy tools, the Yuanbao assistant, and the ima workspace platform. External developers can access the model's application programming interface via Tencent Cloud Tokenhub as well as third-party aggregator OpenRouter.
The release follows an overhaul of Tencent's foundational AI infrastructure earlier in the year. According to the company, the Hunyuan engineering team has maintained an average delivery cadence of one major model version every two months since rebuilding its infrastructure in February.
Sources
- 腾讯混元开源新一代大语言模型Hy4 preview--经济·科技--人民网 — People's Daily · 8/29/2026