Solved Qwen3.8

That is not the same thing as a guaranteed serverless API speed, and AxForge explicitly says it has not published its own tuned number yet.

Real speed changes with prompt length, concurrency, serving engine and batching, so I would not turn 25.1 t/s into an SLA.

Does Qwen3.8 understand images too?

The Qwen3.8 27B model family is vision + text capable.

Whether a specific hosted endpoint exposes every upstream multimodal feature is something I would verify against the current API docs before promising a particular image workflow.

For a FAQ, I would say Qwen3.8 is a vision-language model, but point users to the current endpoint documentation for supported request formats and limits.