Local LLM performance report
Leave unknown or unmeasured values explicitly unknown.

Measurement date and timezone:
Model repository, file/tag and revision/digest:
Total and active parameters, if known:
Weight quantization and format:
GPU model(s), count and VRAM per device:
CPU, system RAM and OS:
Engine version, backend and driver:
Offloaded layers/experts and device split:
KV-cache type and placement:
Configured context limit:
Actual input tokens per run:
Actual output tokens per run:
Concurrent requests:
Batch settings (not the same as request concurrency):
Speculative decoding / draft model:
Other settings and background load:

Exact command/request and prompt (remove private content):
Warmup procedure:
Repetition count:
Cached input / prefix reuse:
Raw output or source URL:

Per-run generation tokens/s:
Per-run uncached prompt-processing tokens/s (or unknown):
Time to first token and measurement method (or unmeasured):
Wall-clock completion time:
Peak memory use and measurement method:
Range/median over repeated runs (never just the fastest):
Loading time included or excluded:
Task quality observations (separate from speed):
Known limitations: