Short-prompt service responsiveness

Warm median completion was 0.891 s on TestBench, 2.345 s on Z13, and 2.080 s on Luna. These are descriptive service observations, not a hardware-only comparison or general speed ranking.

Five matched short prompts

Each service received the same five prompt byte strings in sequence, with a 128-token output cap. Temperature was omitted; Qwen used think:false. No answers were quality-graded.

Service / observed setup	n	TTFT median	TTFT p95*	Completion median	Completion p95*	Output tokens median
TestBench system Ollama / Qwen3.6 35B (resident across two RTX 3090s)	5	0.169 s	0.176 s	0.891 s	0.973 s	77
Z13 Radeon 8050S / Qwen3.6 35B	5	1.111 s	1.118 s	2.345 s	2.482 s	79
Hosted GPT-6 Luna	5	1.478 s	1.840 s	2.080 s	2.648 s	81

* Nearest-rank p95 with n=5; descriptive only. Output token counts differ by service and tokenizer.

What the timings include

The same short prompt bytes were sent for each repeat across lanes; tokenization is model-specific.
Temperature was omitted, so each service used its native/default behavior.
Local Ollama timings are measured on the model host; Luna wall/TTFT includes Pi SDK work from before streamSimple through completion.
Luna's onPayload observation is not wire-confirmed; the Pi Codex builder may inject its default helpful-assistant instruction despite the observed empty instructions field.
The TestBench 64K Pi alias is absent; the existing system Ollama 32K tag was used.
Descriptive service timings only; no quality, hardware-only, or general-ranking claim.

The Z13 model was not resident at the start; its 7.14 s load occurred during the excluded warm-up. The measured requests followed that warm-up. TestBench's pre-existing model remained resident. Thus these are service-level warm observations under their recorded conditions, not a cold-start comparison.

Interpretation boundary

Local Ollama timings were measured on each model host. Luna timing was measured from the desktop and includes Pi SDK work from before streamSimple through completion; it is not provider-side TTFT, prefill, or decode. Luna's observed payload callback is not wire-confirmed, and the Pi Codex builder may supply its default assistant instruction.

This branch answers how quickly these services returned visible text for this small prompt set. It does not measure answer quality, isolate hardware as the cause, or predict long-context behavior. In particular, the earlier 31k-token Z13 stress prompt is not a proxy for this workload.