Tencent Hy4 is available on Vercel AI Gateway, according to Vercel’s November 3 release note. The integration adds a model aimed at coding, reasoning and agentic work to an interface that developers can use from existing applications and coding tools.
The announcement describes image input, reasoning, tool calling and structured outputs, together with a one-million-token context window. It also identifies computer use across command-line and graphical interfaces. Vercel says Hy4 improves on Hy4 Preview in instruction following, long-context work and sustained tasks. Those descriptions come from the provider’s release material.
Available through an existing integration path
Vercel documents the model identifier as tencent/hy4. Its model page provides API examples and information for coding-agent setup, including supported request interfaces. The page lists a 64,000-token maximum output and identifies Tencent Cloud as a provider.
The launch includes a 50% introductory discount through December 31, 2026. Teams comparing costs should retain that end date in their calculations. A promotional rate can make an evaluation economical without establishing the long-term cost of a production workload.
Vision and computer use need different checks
Our analysis is that Hy4’s combination of capabilities is most interesting where a task crosses representations. A coding assistant may need to understand a screenshot, change source files and inspect the resulting application. A document assistant may need to read a chart and return structured information that another program can validate.
Those tasks require more than a model option. Computer use depends on the surrounding application supplying an execution environment, observations and permission controls. Selecting a model that supports the capability does not automatically give an assistant access to a browser or authorize changes to external accounts.
A sensible initial trial would use a disposable environment and a task with a clear result. Ask the assistant to reproduce an interface from a reference, run the project and check the page. Compare the completed result, elapsed time and total inference spend with the current model using the same environment and instructions.
Structured outputs deserve a separate test with the actual schema the product expects. Include missing fields and ambiguous inputs, then verify that the application rejects malformed or unsupported values. This reveals whether the integration is dependable under ordinary uncertainty.
Hy4’s arrival expands the set of models accessible through a common gateway. The practical value will come from how well it handles a team’s complete workflows, especially those that combine visual input, executable actions and a verifiable final result.