Tencent Hy4 preview is a large mixture-of-experts language model that activates 49B of its 770B total parameters per request, paired with a context window exceeding one million tokens that enables very long document, codebase, or transcript handling without external retrieval tricks. Tencent describes the release as a major step up in model size, context length, and training data volume over its predecessor, with the model co-designed alongside Tencent's own productivity products so that feedback from real coding and office workflows informed the training and post-training process. The same Hunyuan lineage that produced earlier Hy generations therefore underpins both the architectural choices and the practical emphasis on day-to-day knowledge work rather than purely academic benchmarks.
Independent leaderboards place Hy4 preview near the top of the open-weight field, sitting 11th overall on the LLM Stats composite score and landing in the top ten percent for tool calling, reasoning, and coding, while long-context behavior is described as average and specialized domains like legal and finance remain weaker relative to peers. Tencent highlights strong results on economically valuable knowledge-work tasks such as GDPval and expert-level graduate questions like GPQA Diamond, alongside its internal blind evaluation with 163 domain experts, reinforcing a positioning aimed at developers, analysts, and researchers who need a self-hostable model for substantial coding sessions, document-heavy analysis, or scientific writing with iterative tool use.
