Tencent’s Hunyuan team has published Hy4-preview on Hugging Face under an Apache 2.0 license. The mixture-of-experts model is listed at 770 billion total parameters with about 49 billion activated per token, 256 routed experts plus one shared expert, and a claimed 1-million-token context window.
The model card cites 92.3 on GPQA Diamond and 65.7 on SWE-bench Pro. Weights are offered in FP8, with Docker recipes for vLLM and SGLang. Tencent is positioning Hy4 against other Chinese frontier systems from Z.AI and Moonshot, part of a broader wave of large open or semi-open releases that Bloomberg says is narrowing the U.S.–China capability gap on price if not always on closed-lab peak scores.
A 1-million-token window at this scale matters for long-document legal work, multi-repo coding, and agent traces that used to require retrieval hacks. Apache 2.0 plus published serving recipes also lowers the barrier for labs that want to fine-tune or host the model themselves rather than buy API tokens from U.S. vendors.
Independent third-party evals are still catching up to the vendor card. As with other mega-MoE releases, the practical questions are serving cost at 49B active parameters, quality at the far end of the context window, and whether export and chip-supply constraints limit who can actually run Hy4 at full size.
Key takeaway — Tencent’s Hy4-preview puts a 770B MoE model with a 1M-token context and Apache 2.0 weights on Hugging Face, adding another large, cheap-to-license Chinese system to the open-weight race.
Photo: Google DeepMind / Unsplash · Sources: TechNode; AI Weekly; Bloomberg.
