Tencent launches Hy4 with image understanding Tencent launches Hy4 with image understanding

Hosted access comes first

Tencent reports computer use through command-line and graphical interfaces, with targeted training on Blender and CAD. The company describes applications in chart analysis and visual mathematics. Hy4 is available through TokenHub, WorkBuddy and CodeBuddy, with API access through platforms including OpenRouter.

The context ceiling is unchanged

Tencent Cloud’s catalog lists separate hy4 and hy4-preview identifiers. Both have a one-million-token context window, a 960k-token maximum input and a 64k-token maximum output. Both entries retain structured output, function calling, caching and preserved thinking. The published input and output caps are separate limits, not permission to fill the entire context with input. The catalog does not list video understanding for Hy4.

A temporary discount changes the bill

Tencent Cloud’s Chinese-language pricing table lists RMB 6 per million input tokens, RMB 18 per million output tokens and RMB 0.30 per million cache-hit tokens. These are the same list rates shown for the preview. Hy4 usage is billed at half the listed price through December 31 at 23:59:59 Beijing time, then returns to the standard rate. Applying that discount gives RMB 3, RMB 9 and RMB 0.15 respectively. A hypothetical workload of one million uncached input tokens and one million output tokens therefore costs RMB 12 during the promotion, versus RMB 24 afterward. That calculation excludes cache hits and describes token charges, not a measured task-completion cost.

Image delivery has its own limits

The Hy4 integration guide supports image input through Chat Completions, Responses and Anthropic-compatible protocols. Images can be mixed with text across multiple turns. Each downloaded or Base64-decoded image file may be at most ten megabytes; a request may include six hundred images, counting conversation history. The detail setting also matters. Low detail scales an image within 512 by 512 pixels, while the original setting is available only through Responses. The guide recommends higher detail for small text, tables and screenshots. It says images ultimately scale within roughly 2048 by 2048 pixels, so sending a larger file does not expose additional detail to the model. Inputs exceeding the documented 5500-by-5500 resolution boundary are rejected. Supported formats are PNG, JPEG, WEBP and non-animated GIF. Historical image messages must be resent on subsequent turns; the service downloads URL images or decodes Base64 again on each request. For streamed Chat Completions calls, include_usage must be enabled to receive complete usage statistics in the final chunk.