RES — Latency on the reference laptop profile is consistently under 3 s (P95)…
Latency on the reference laptop profile is consistently under 3 s (P95) for small models (⇐3B) and over 5 s (P95) for large models (>=8B); one 7B model showed extreme variability (10x-18x) from thermal throttling and scheduling, unsuitable for strict latency SLAs.
Source: PUMA project documentation · Traceability: corpus unit
MEM-052· Confidence: verified-at-primary-source