Tencent announced the official Hy4 release on November 4, adding visual understanding and capabilities for working with command-line and graphical interfaces. It lists access through its applications and cloud services, while promising model weights and system cards in the following days. The launch announcement therefore should not be treated as evidence that every file needed for self-hosting is already available.
From Preview to the official release
Hy4 Preview was introduced on August 28. At that time Tencent described 770 billion total parameters, 49 billion active parameters and a context exceeding one million tokens. Those figures belong to Preview and should not automatically be transferred to the new release. The earlier model was already available in WorkBuddy, CodeBuddy and other company products.
That history matters because the name of a model and the name of a service often appear together but identify different things. Replacing a model inside an application can change an existing workflow without the user installing a new program. This is editorial context for evaluating an update, not a claim that any particular customer system was changed without notice.
What the cloud documentation confirms
The Tencent Cloud ADP team confirms that Hy4 entered its workspace on November 4. It describes a service with approximately 1.05 million context tokens and up to 65,536 output tokens, with exact limits depending on the access route. The post discusses configuring permissions, evaluating and publishing agents, and inspecting execution records. A temporary discount runs through December 31, 2026; it is not a permanent price for every possible service.
We have not independently benchmarked the model. Provider capability claims and rankings should be distinguished from a team’s own tests on its actual tasks.
What to check in a real workflow
Our editorial recommendation is to start evaluation with a completed task: for example, whether an agent can read a diagram, make the correct change and save a usable result. A convincing explanation of an image and successful operation of an application are different outcomes. The latter also depends on permissions, environment state and a person’s ability to review what was done.
A useful comparison would keep the task, tool access and completion criteria consistent between versions. Otherwise a better result might reflect a different prompt or a more capable application wrapper rather than the model alone. This is a suggested evaluation approach, not an additional benchmark result reported by Tencent.
Likewise, a long context window describes input capacity, not a guarantee that every detail will be used correctly. Teams considering self-hosting should inspect the files and terms actually released rather than rely on the word “open” in a headline. The established news is the new version and confirmed cloud access, with additional artifacts described as a subsequent step.
