Hy4 preview

What is Hy4 preview?

Hy4 preview is Tencent’s text model for coding agents and workflows that require planning across several steps. OpenRouter exposes it as a paid model with a large context window. Consider a controlled repository task with a clear stopping condition; this preview needs an application or agent framework to perform actions.

Plan before changing a repository

Tencent’s model card positions Hy4 preview around productivity work, including long-horizon software engineering: planning, debugging and verifying changes. A practical assignment is to trace a feature through a small codebase: provide an entry point and ask which files participate, what assumptions connect them and where a proposed change should be tested.

For a recurring operational task, give it a written procedure and a sample log. Ask for the next step and the evidence that justifies it. Treat the result as a proposal until the underlying command, destination and expected effect are checked.

Suggested workflow: trace one feature

Choose a feature with a known input and output. Supply relevant files and a brief architecture note, then ask for an execution path with file references. Require the model to identify missing dependencies rather than guessing their contents.

Check each reference against the repository. After approving the map, request one narrow change and the checks that should pass. Compare the patch with the plan, run those checks and stop if the model starts editing unrelated areas. Save the result as a repeatable evaluation case.

Preview access and provider pricing

Use tencent/hy4-preview through OpenRouter, which bills hosted usage per token. Before 16:00 UTC all four listed providers charged $0.834 per million input tokens and $2.501 per million output tokens; from 16:00 to 24:00 UTC the Tencent endpoint lists rates 10% lower, so any single price is only valid for its date.

Before a larger run, inspect the current endpoint, cap the task and measure its actual usage. Tencent publishes a model card and deployment resources; self-hosting is a separate operational choice, not a feature included with an OpenRouter request.

When to choose a different input path

For a UI problem that depends on a screenshot, consider a multimodal model such as GLM 5.3 Flash instead, or convert the visual findings into an accurate written brief first. Hy4 preview’s large text window does not turn it into an image-reading model.