Tencent Hy4 Preview is a next-generation open-source large language model designed for real-world productivity work spanning coding, office tasks, and scientific research. It uses a 770B-parameter Mixture-of-Experts design that activates 49B parameters per inference, combining a very large knowledge base with efficient per-query compute. The model was scaled up jointly in size, context length, and training data volume, with Tencent describing coordinated advances in both pre-training and post-training that produced a step-change in overall intelligence. A context window exceeding one million tokens lets it reason over very long documents, code repositories, or multi-step research workflows without losing earlier context.
Practical deployment is a focus: Tencent co-designed the model with its own productivity products such as CodeBuddy and WorkBuddy, so the training and evaluation pipeline was tuned toward the kinds of tasks these tools actually field. An internal blind evaluation with 163 experts and 203 engineering tasks was used to measure real-world performance, signaling that engineering productivity — not just academic benchmarks — drives its development. Tencent positions it among the top tier of open-source models, though this ranking is currently a vendor self-claim until independent benchmarks corroborate it. The combination of open weights, very long context, MoE efficiency, and product-grounded training makes Hy4 Preview a strong fit for teams that want a self-hostable base model for code agents, long-document analysis, and scientific assistants.
