Tencent releases open-weight Hy3 model
Tencent has released the final version of Hy3, a 295B-parameter open-weight MoE model under Apache 2.0, which already runs across its consumer and enterprise products and claims to match much larger flagships.
Tencent announced the official release of Hy3, the full version of the model it previewed on 23 April. Hy3 is a hybrid fast-and-slow-thinking Mixture-of-Experts model with 295 billion total parameters, 21 billion active parameters, and a context window of up to 256K tokens.
The weights are available under the Apache 2.0 license on Hugging Face and ModelScope. The API is live on Tencent Cloud TokenHub. Tencent says it will roll Hy3 out gradually to third-party platforms including OpenRouter, Kilo, Cline, OpenCode and Cherry Studio.
Tencent says Hy3 reaches intelligence "comparable to flagship models with two to five times its parameter scale," and is more stable and cheaper to run than the preview. It credits the gains to stronger reinforcement learning and better training data. It names software development, office work, financial modeling, front-end design and game production as areas of improvement. The announcement includes no benchmark figures to back up these claims.
Inside Tencent, Hy3 now powers WorkBuddy/CodeBuddy, the Yuanbao assistant (which has added a free Agent feature that creates PowerPoint, Word, Excel, PDF and HTML files), the ima knowledge base, the OS-level Marvis Agent, and customer-service assistants on Weixin Official Accounts. Tencent says average daily token consumption has grown twentyfold since the preview launched. It also says the number of WorkBuddy users who choose the model has grown sixfold.
Tencent, which now calls its Hunyuan models "Tencent Hy," says it rebuilt its model infrastructure in late January 2026. That puts the time from rebuild to this release at under six months.