Overview
On November 4, Tencent Hunyuan released Hy4, its new flagship large language model. Versus the earlier Hy4 preview, pre-training and post-training data and scale were significantly expanded: total parameters reach 770 billion (770B) using a Mixture-of-Experts (MoE) architecture (the Hy3 series was under 300B). Hy4 introduces vision understanding for the first time and, notably, Computer Use — operating both command line and graphical interfaces to handle professional software such as Blender and CAD and cross-application long-horizon tasks.
The company said it will soon open-source the model weights, publish a model card, and focus on web and biological-safety risks; a Hy4 Flash variant will follow. Integrated with the WeCom suite, Hy4 can connect chats, mail, documents, schedules and meetings to identify follow-up tasks and upcoming deadlines.
Background & Interpretation
1. Technical Background and Evolution
The competitive focus is shifting from answer accuracy to whether the model can operate software for people. Hy4’s jump from under 300B to 770B while pledging open weights effectively transfers the cost of super-large MoE plus engineering deployment to cloud providers and inference optimizers, further eroding the scale moat of closed models. Vision plus Computer Use signals domestic models entering a perceive—decide—act integrated stage.
2. Core Drivers and Underlying Mechanism
The MoE architecture scales parameters while controlling inference cost via sparse activation, making 770B-class models usable. Computer Use relies on joint modeling of GUI and CLI, requiring long-horizon planning and fault tolerance on real interfaces. Tencent’s WeCom ecosystem provides a natural Agent entry point, embedding capability into collaboration workflows.
3. Market Structure and Industry Chain Impact
The parameter leap plus open-source commitment intensifies the domestic scale + openness dual competition, forcing cloud providers to optimize inference cost. For the application layer, Computer Use lowers the bar for office automation and professional-software operation, likely spawning GUI-automation agents. Compute, inference optimization and model safety evaluation become pivotal.
Implications & Outlook
Domestic models are collectively moving toward larger parameters + multimodality + executability, with agentification and open-sourcing in parallel. Watch: actual release pace of Hy4 weights and card, Computer Use task-completion and fault-tolerance on real UIs, and the inference-cost decline curve.
Takeaways for Industry Participants
Evaluate such models by long-horizon completion rate and per-task cost before benchmarks. Enterprises can pilot in office automation and professional-software operation. Developers should seize fine-tuning and private-deployment chances from open weights while respecting web and bio-safety compliance boundaries.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.