Tencent’s Hy3 MoE preview surfaces

Tencent said on April 23 that it released and open-sourced Hy3 preview, a new Hunyuan language model built around a mixture-of-experts architecture. (hunyuan.tencent.com) (cloud.tencent.com) A mixture-of-experts model works like a team of specialists: the full system is large, but each prompt activates only part of it. Tencent says Hy3 preview has 295 billion total parameters and 21 billion active parameters. (cloud.tencent.com) Tencent says the model supports up to a 256K context window, which is the amount of text it can keep in view at once. The company also says Hy3 preview is aimed at code generation, agent workflows, complex reasoning, and long-document instruction following. (hunyuan.tencent.com) (cloud.tencent.com) Tencent framed Hy3 preview as the first model to come out of a broader rebuild of its Hunyuan training stack that began in February 2026. The company’s research site labels it “the first step in rebuilding the Hy model.” (hunyuan.tencent.com) Tencent says that rebuild covered both pretraining and reinforcement learning, the post-training method used to shape how a model answers. In Tencent’s telling, the goal was less about headline parameter counts and more about lower serving cost and steadier performance in product use. (cloud.tencent.com) The company tied Hy3 preview to internal benchmarks and product tests rather than only public leaderboards. Tencent says the model posted strong results on SWE-Bench Verified, Terminal-Bench 2.0, BrowseComp, and WideSearch, though the source surfaced here is Tencent’s own description rather than an independent benchmark report. (cloud.tencent.com) Tencent also says Hy3 preview had already been gray-released inside products including CodeBuddy, WorkBuddy, Yuanbao, ima, Peacekeeper Elite non-player characters, and Tencent Docs AI PowerPoint tools. In those tests, Tencent reports first-token latency fell 54% and end-to-end time fell 47% in its developer tools. (cloud.tencent.com) On pricing, Tencent says Hy3 preview is available through Tencent Cloud TokenHub at 1.2 yuan per million input tokens, 0.4 yuan per million cached input tokens, and 4.0 yuan per million output tokens. Tencent also says it is offering a Hy3 preview token plan priced at 28 yuan a month for individual developers. (cloud.tencent.com) Tencent’s own model catalog shows the company has been expanding a wider Hunyuan lineup that includes other long-context and agent-focused systems, including Hunyuan-A13B and the newer HY 2.0 series. Hy3 preview fits that same push toward models that trade full activation for lower inference cost. (hunyuan.tencent.com 1) (hunyuan.tencent.com 2) The immediate news is narrower than the social-media preview suggested: Tencent did not just tease Hy3 preview, it said the model was released and open-sourced on April 23. That turns the story from a rumor about specs into a formal product and developer launch. (hunyuan.tencent.com) (cloud.tencent.com)