DeepSeek V4.1 Flash 552B (16B active)

User reports DeepSeek V4.1 Flash at 20.1 tok/s on an unnamed machine, served through Maya. The run read 223 new prompt tokens in 3.9 s with 65 reused, over a 1,155-token conversation; no quant, context length or KV cache setting is given. The user's point is behavioural: the model insisted it is Claude and refused to repeat its prompt verbatim, and the user asks whether Maya diverges from upstream in the scheduler or offload path.