Tencent Open-Sources Hy4 Preview, a 770B-Parameter Model With 1M Token Context

Tencent has released and open-sourced Hy4 preview, a next-generation large language model designed for enterprise productivity. The release puts a 770-billion-parameter system with an unprecedented 1-million-token context window into developers' hands.

The model features 770 billion total parameters with 49 billion active during inference, according to Tencent's announcement. Its context window exceeds 1 million tokens, enabling processing of extremely long documents and codebases. In an internal blind evaluation conducted by 163 experts across 203 engineering tasks, Hy4 preview scored 2.99 out of 4.00, outperforming both GLM-5.3 at 2.92 and Kimi K3 at 2.94. The model is already integrated into Tencent's productivity suite, including WorkBuddy, CodeBuddy, and Yuanbao. API access is available through Tencent Cloud TokenHub and OpenRouter, with pricing set at USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million tokens for cache hits.

A Productivity-First Architecture

Tencent is positioning the Hy4 series as a productivity-focused model, co-designed with its internal tools and trained on high-quality data from experts in software engineering, gaming, finance, and security. The company is employing a "preview-first" strategy, releasing the model to gather real-world feedback before an official full release. This approach allows Tencent to incorporate user insights and identify potential improvements based on actual deployment scenarios. The hybrid parameter design—770 billion total with only 49 billion active—suggests a mixture-of-experts architecture optimized for efficiency while maintaining capacity for complex reasoning tasks.

Competitive Implications for Open AI

The release pushes boundaries for open-weights AI in enterprise contexts, particularly for tasks requiring extensive context retention. At 770 billion parameters, Hy4 preview ranks among the largest openly available models, competing directly with proprietary systems from major AI labs. The pricing structure undercuts many closed alternatives while offering the flexibility of open-source deployment. For enterprises evaluating AI infrastructure, the combination of massive context windows and competitive pricing could shift procurement decisions toward open-weight solutions. However, the internal evaluation metrics, while promising, have not been independently validated by third-party researchers.

What Developers Should Watch

The model's actual performance in production environments, beyond internal evaluations, will determine whether Hy4 preview delivers on its productivity promises. Developers and enterprises will be watching closely to see how the model performs across diverse real-world workloads. Key questions include whether the 1-million-token context window maintains accuracy at extreme lengths, and how the model handles domain-specific tasks in finance, security, and gaming where Tencent claims specialized training. The preview release strategy suggests Tencent expects iterative improvements based on community feedback before committing to a final version. Independent benchmarking will be crucial for validating Tencent's performance claims against competing models. Early adopters should document their findings and share results with the broader community to establish reliable performance baselines.