Key highlights Tencent Hunyuan has released its new flagship model, Hy4 preview, and the headline numbers are striking: 770B total parameters, 49B active parameters, and a context length of 1M tokens. Just as important, the model is open-source its weights are freely available for use and it has been launched simultaneously on Tencent Cloud's model repository and on OpenRouter, the aggregator that exposes many models through a single API. For China's open-source large-model camp, this is a double refresh of both parameter scale and context length, and a clear signal that the domestic open-weight race is accelerating rather than slowing down as more players enter the field and raise the bar for what counts as competitive among peers. ## What happened Hunyuan did not follow the rhythm of commercializing behind closed doors first and then gradually opening up. Instead, it released open weights directly in a preview form, inviting the community to test and build early rather than waiting for a polished final cut that might arrive months later. Developers can either pull the model from Tencent Cloud's model repository or call it through aggregation platforms like OpenRouter via API. In other words, both private self-hosting and cloud-based calling are unblocked at the same time, lowering the onboarding cost for users with different needs and different operational constraints. This open-source-and-immediately-usable release style also lets the community obtain the base model faster for secondary development, whether that means fine-tuning, distillation, or building downstream applications on top without waiting for an official partner program to open. ## Technical details The pair of numbers 770B total and 49B active parameters essentially points to a Mixture-of-Experts (MoE) architecture. In such a design the model contains a large number of expert sub-networks, but each inference step activates only about 49B of them. The benefit is that total capacity reaches the 770B level while the per-token compute cost stays close to that of a 49B dense model, balancing capability against cost in a way dense models cannot match. The 1M context length, meanwhile, means the model can ingest roughly a million tokens in a single window an entire book, a whole codebase, or an extremely long meeting transcript which is highly valuable for long-document understanding and for agentic tasks that must keep a great deal of state in memory at once while reasoning across distant parts of the input without losing the thread that connects them. ## How it compares Zooming out to the open-source camp, million-level context windows together with hundreds-of-billions-parameter MoE models have, over the past year, become a standard direction for leading labs around the world. Hunyuan Hy4 preview's 1M context sits in the top tier on paper, but what really matters in practice is the effective utilization under long context. A long window does not by itself guarantee that every piece of information receives fair attention; performance at the far end of the context often degrades, and that usually requires both clever architecture and careful training data to ensure. Open weights let the community run its own stress tests and reproductions, which can actually speak louder than closed-source leaderboards and more easily surface degradation or omissions hiding inside long context that a vendor-controlled benchmark might miss because it was tuned to flatter the model. ## Why it matters In plain terms, the open-sourcing of Hy4 preview means more than yet another downloadable model. It puts the latest domestic capabilities in long context and large-parameter MoE into developers' hands in a self-hostable form. For enterprises that need data to stay within their own boundaries, teams doing long-document processing, and researchers who want to fine-tune on a flagship base, this is a concrete and attractive option. Whether the experience ultimately lives up to the 1M-window promise will depend on real community evaluations once it goes live, and open weights are precisely what make such independent, large-scale evaluations possible in the first place, rather than leaving the claims to be judged only by the provider who has an incentive to present them in the best light. The practical upshot is that the model's reputation will be built by the community, not the marketing team, which tends to produce a more honest picture of strengths and weaknesses over time. For developers in China especially, having a domestically produced flagship that is both open and competitive on paper reduces dependence on foreign-hosted models and the geopolitical risk that comes with routing sensitive data through them. Because the release carries a preview label, the weights are also likely to improve through subsequent iterations, and the open form means those improvements can be adopted without renegotiating a vendor relationship. The coming weeks of community testing will be worth watching closely by anyone building on open-weight foundations. If the community testing confirms the long-context claims hold up in practice, Hy4 preview could become a default choice for teams that need both scale and sovereignty without leaving the open-weight ecosystem.
