
Tencent released Hy4 Flash on November 16, adding a cheaper model alongside flagship Hy4. It supports a one-million-token context window and targets document processing, research, simple front-end tasks and computer use. The official announcement positions Flash for everyday workflows.
Standard prices, separate from launch offers
Tencent quotes approximate standard API prices of $0.070 per million input tokens and $0.281 per million output tokens, with cached input at $0.003 per million tokens. Temporary launch discounts are separate from those rates.
Access is available through Tencent Cloud TokenHub and WorkBuddy. The launch announcement describes OpenRouter integration as coming soon, rather than an already available route.
Visual inputs connected to tools
TokenHub's model listing specifies text and image input, text output, function calling and structured outputs—the ingredients for connecting visual information to software tools. The official model catalog also lists the hy4-flash API identifier and its supported capabilities.
Why it matters for builders
For small teams such as SunMarc App Labs, useful evaluation targets include:
- Extracting structured information from screenshots and documents.
- Routing routine tasks to a cheaper model while reserving stronger models for harder work.
- Comparing cost per successful task, including retries and human corrections—not token prices alone.
The evaluation should use real examples with known answers. A low token bill is not useful if missing fields or incorrect tool actions create more review work. We have not tested Hy4 Flash; these are proposed experiments, not measured savings.
Keep the flagship and weights distinctions clear
Tencent still recommends flagship Hy4 for complex, long-running software engineering. Open-sourcing Flash is planned; that is not confirmation of downloadable weights.
For the earlier model and its visual-agent positioning, see our flagship Hy4 coverage. Flash's lower price is a reason to evaluate task routing, not evidence that it should replace a stronger model across every workflow.