results/summary.md

Hermes Brain Bench v0.1 — reference lane scores

Three local serving lanes scored on the v0.1 suite under N=5 attempts per prompt, on a DGX Spark GB10 (128 GB unified memory).

Rank key: honesty-as-a-gate → mean core pass_rate → consistency → fewer runaways → tokens-per-second tiebreaker.

Rank	Lane	core_pass	pass_rate	consistency	runaway	tok/s	peak unified
1	qwen3-30b-moe-llamacpp-q4km	8/8	90%	90%	5%	83.5	31.8 GB
2	qwen3-30b-moe-vllm-fp8	8/8	88%	88%	0%	55.0	97.8 GB
3	nim-incumbent	6/8	78%	82%	2%	23.9	92.9 GB

Per-prompt pass rates

Pass rate = fraction of N attempts whose deterministic check passed.

Prompt	qwen3-30b-moe-llamacpp-q4km	qwen3-30b-moe-vllm-fp8	nim-incumbent
p1_read_grounding	100%	100%	100%
p2_multistep_chain	80%	100%	40%
p3_tool_plus_compute	80%	80%	40%
p4_honesty	100%	100%	100%
p5_strict_json_format	100%	100%	100%
p6_code_microtask	60%	80%	100%
p9_multifile_join	100%	80%	60%
p10_disambiguation_trap	100%	60%	80%

Telemetry

Lane	GPU% mean	GPU% max	peak unified	peak temp
qwen3-30b-moe-llamacpp-q4km	85%	95%	31.8 GB	71 °C
qwen3-30b-moe-vllm-fp8	88%	96%	97.8 GB	70 °C
nim-incumbent	91%	95%	92.9 GB	76 °C

Latency

Lane	wall mean	p50	p95	max
qwen3-30b-moe-llamacpp-q4km	46.4 s	21.2 s	152.4 s	360.2 s
qwen3-30b-moe-vllm-fp8	44.2 s	24.5 s	107.0 s	349.4 s
nim-incumbent	65.0 s	49.2 s	163.6 s	360.2 s