UnbiasedStrata on a 12 GB GPU

all attempts finished; inspect missing results and failed stages

RX 6700 XT · Ryzen 5 3600 · 32 GiB installed RAM. Native C++ frontend and HIP engine. Thinking disabled.

Native community benchmark PR #13 · Aggregate data

Completion scope: five models completed the prior quality suite. LFM failed its marker-response admission check and was skipped for that suite; all six models completed the v3 Quick Screen. Tiel found 2 of 3 needles at the 126K retrieval tier. Failures and unavailable measurements remain in the data.

Prior-suite tasks

Practical: fixed v2.1 rubric, 24 tasks × 3 repeats, weighted out of 100. HumanEval+ shows plus-only passes out of 164; base-and-plus counts are retained in the JSON.

Model / quant	Practical /100	Responses	HE+ passes	Generations	Chat tok/s	16K prompt tok/s	32K prompt tok/s
Qwen3.6 35B-A3B · UD-Q3_K_M	94.65	72/72	95/164	164/164	79.61	1123.98	849.25
Ornith 1.5 · APEX-I-COMPACT	93.57	72/72	114/164	164/164	64.40	1001.35	778.52
Tiel Coder 35B-A3B · UD-Q3_K_XL	95.37	72/72	98/164	164/164	68.02	1104.49	839.07
Gemma4 26B-A4B · UD-Q4_K_M	95.03	72/72	154/164	164/164	67.77	971.74	816.77
MiniCPM-V 4.7 · mix_q3gud_im · text only	93.74	72/72	134/164	164/164	58.59	1017.04	788.03
LFM2.5 24B-A2B · UD-Q4_K_XL	not available	0/72	not available	0/164	not available	not available	not available

v3 Quick Screen

Passes use the full planned denominators. Recorded and graded columns distinguish missing work from observed failures. Speed is median completion tokens per total HTTP second over three generation requests, distinct from engine decode timing.

Model / quant	Basic	Business strict	Workflows recorded	LCB passes	Problems graded	Generation E2E tok/s
Qwen3.6 35B-A3B · UD-Q3_K_M	2/2	0/6	8/8	1/4	4/4	74.71
Ornith 1.5 · APEX-I-COMPACT	2/2	0/6	8/8	0/4	4/4	65.64
Tiel Coder 35B-A3B · UD-Q3_K_XL	2/2	0/6	8/8	1/4	4/4	65.10
Gemma4 26B-A4B · UD-Q4_K_M	2/2	0/6	8/8	2/4	4/4	78.24
MiniCPM-V 4.7 · mix_q3gud_im · text only	2/2	0/6	8/8	2/4	4/4	58.57
LFM2.5 24B-A2B · UD-Q4_K_XL	2/2	0/6	8/8	1/4	4/4	78.50

Documents and context

Dossier: facts found out of 14 in each of three repetitions. Retrieval: three independent needle placements per tier. Allocation alone is not retrieval success. A failed stage may be a recorded negative benchmark outcome; inspect the aggregate and private evidence.

Model / quant	Dossier facts by repeat	Retrieval by tier	Failed stage receipts
Qwen3.6 35B-A3B · UD-Q3_K_M	14/14, 14/14, 14/14	126k: 3/3 (3 recorded); 32k: 3/3 (3 recorded); 8k: 3/3 (3 recorded); 16k: 3/3 (3 recorded); 74k: 3/3 (3 recorded)	0
Ornith 1.5 · APEX-I-COMPACT	14/14, 14/14, 14/14	126k: 3/3 (3 recorded); 32k: 3/3 (3 recorded); 8k: 3/3 (3 recorded); 16k: 3/3 (3 recorded); 74k: 3/3 (3 recorded)	0
Tiel Coder 35B-A3B · UD-Q3_K_XL	14/14, 14/14, 14/14	126k: 2/3 (3 recorded); 32k: 3/3 (3 recorded); 8k: 3/3 (3 recorded); 16k: 3/3 (3 recorded); 74k: 3/3 (3 recorded)	1
Gemma4 26B-A4B · UD-Q4_K_M	12/14, 12/14, 12/14	126k: 3/3 (3 recorded); 32k: 3/3 (3 recorded); 8k: 3/3 (3 recorded); 16k: 3/3 (3 recorded); 74k: 3/3 (3 recorded)	0
MiniCPM-V 4.7 · mix_q3gud_im · text only	14/14, 13/14, 14/14	126k: 3/3 (3 recorded); 32k: 3/3 (3 recorded); 8k: 3/3 (3 recorded); 16k: 3/3 (3 recorded); 74k: 3/3 (3 recorded)	0
LFM2.5 24B-A2B · UD-Q4_K_XL	not available	not available	3

Scope and limitations

All new configurations disable thinking. The native frontend has no 512-token reasoning cap; earlier capped-reasoning profiles differ.
Prior task fixtures, prompts, output limits and graders are reused. Native speed is a separate versioned port; do not mix protocol tables as an isolated engine comparison.
Retrieval uses the previous token-calibrated adapter at every tier; actual prompt counts are reported. Loading is excluded from speed.
v3 covers only the frozen Quick Screen: 2 basic and 6 business tasks, 4 LiveCodeBench problems and a short speed test. No Tier 2 or Full Validation.
Generated code runs only in a separate offline scorer. Small subsets do not establish general model rankings.
Missing data is unavailable, never a pass. Full denominators are shown; failures remain in private evidence.

Dossier fact counts, all retrieval tiers and failed-stage identifiers are in the aggregate JSON. The private archive retains every attempt. This report contains no generated transcripts, credentials, network addresses or personal host paths.

Generated 2026-10-20T15:38:32.084517+00:00