Model and quant benchmarks

Loading benchmark snapshots…

2026-10-26T08:25:01.000Z · Configured data sources · viewer

Models needing benchmark runs

Active or planned models that are not complete against the standard capability suite.
Retired models are excluded from this queue.

Priority model	Status	Coverage	Missing benchmark families	Last run
Gemma-4-26B-A4B-it	planned	0/13	sql, review, sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	not started
Nanbeige4.3-33B	planned	0/13	sql, review, sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	not started
Qwen3.6-35B-A3B	planned	0/13	sql, review, sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	not started
Claude Fable 5.1	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T01:19:24.648519292Z
Claude Opus 5.5	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T01:15:26.872044115Z
Claude Sonnet 5.5	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T01:11:22.303152617Z
Gemma-4-12b-it	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T23:03:58.394274969Z
Gemma-4-31B-it	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T01:09:24.512369098Z
GPT 5.6 Terra	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T01:38:41.405459879Z
GPT 6 Astra	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T01:33:04.115683002Z
GPT 6 Luna	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T01:28:10.213612225Z
GPT 6.1 Sol	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T01:24:42.289154738Z
LFM2.5-24B-A2B	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T05:33:37.886905249Z
LFM2.5-8B-A1B	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T07:54:48.294699165Z
Qwen3.6-27B	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T20:59:17.363160309Z
Qwen3.8-27B	active	2/13	sqlhard, codehard, swe, arch, perf, testgen, gui, coding, logic, science, msms	2026-10-25T22:51:54.660018136Z

Model and quant benchmarks

Models, quantizations and runtime placements across versioned benchmark tasks. Compare observations within their declared hardware, context and measurement source.

Read-only snapshots · refresh every minute while visible

Content-free export; private host labels omitted. Exported speed groups do not establish exact fingerprint/context comparability or qualification.

Overview

16models and configurations

2629reported observations

16incomplete models

4reported lanes

7observed placements

Observation counts are not unique tasks. Missing measurements are shown as —. No inference of qualification or hardware fit.

Published observations retain model, hardware, runtime, card count and prompt context. Runtime version, KV precision, concurrency, cache state and exact deployment fingerprints are not exported; these observations are not equivalent comparison groups.

Decode speed observations

tokens/s · first 24 of 90 rows; all values in the table

Chart data and comparison dimensions

Decode speed observations — exact data
Group	Value (tokens/s)	Dimensions and source
LFM2.5-24B-A2B Q4_K_M KV f16 · 39 ctx	227.7	LFM2.5-24B-A2B Q4_K_M KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 39 prompt tokens · fixed-0 · total 1.4 · TTFT 0.26s · measured
LFM2.5-24B-A2B Q4_K_M KV f16 · 39 ctx	214.1	LFM2.5-24B-A2B Q4_K_M KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 39 prompt tokens · fixed-1 · total 1.2 · TTFT 0.03s · measured
LFM2.5-24B-A2B Q4_K_M KV f16 · 39 ctx	224.9	LFM2.5-24B-A2B Q4_K_M KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 39 prompt tokens · fixed-2 · total 1.2 · TTFT 0.03s · measured
LFM2.5-24B-A2B Q4_K_M KV f16 · 3941 ctx	216.4	LFM2.5-24B-A2B Q4_K_M KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3941 prompt tokens · 4k-cold · total 1.6 · TTFT 0.38s · measured
LFM2.5-24B-A2B Q4_K_M KV f16 · 3941 ctx	215.6	LFM2.5-24B-A2B Q4_K_M KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3941 prompt tokens · 4k-cached · total 1.2 · TTFT 0.02s · measured
LFM2.5-24B-A2B Q4_K_M KV f16 · 15637 ctx	209.1	LFM2.5-24B-A2B Q4_K_M KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cold · total 2.7 · TTFT 1.46s · measured
LFM2.5-24B-A2B Q4_K_M KV f16 · 15637 ctx	207.4	LFM2.5-24B-A2B Q4_K_M KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cached · total 1.3 · TTFT 0.03s · measured
LFM2.5-24B-A2B Q4_K_M KV f16 · 29425 ctx	194.6	LFM2.5-24B-A2B Q4_K_M KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29425 prompt tokens · 30k-cold · total 4.2 · TTFT 2.84s · measured
LFM2.5-24B-A2B Q4_K_M KV f16 · 29425 ctx	190.6	LFM2.5-24B-A2B Q4_K_M KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29425 prompt tokens · 30k-cached · total 1.4 · TTFT 0.05s · measured
LFM2.5-24B-A2B Q4_K_M KV q8_0 · 39 ctx	213.3	LFM2.5-24B-A2B Q4_K_M KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 39 prompt tokens · fixed-0 · total 1.5 · TTFT 0.29s · measured
LFM2.5-24B-A2B Q4_K_M KV q8_0 · 39 ctx	199.9	LFM2.5-24B-A2B Q4_K_M KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 39 prompt tokens · fixed-1 · total 1.3 · TTFT 0.03s · measured
LFM2.5-24B-A2B Q4_K_M KV q8_0 · 39 ctx	199.7	LFM2.5-24B-A2B Q4_K_M KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 39 prompt tokens · fixed-2 · total 1.3 · TTFT 0.02s · measured
LFM2.5-24B-A2B Q4_K_M KV q8_0 · 3941 ctx	196.2	LFM2.5-24B-A2B Q4_K_M KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3941 prompt tokens · 4k-cold · total 1.7 · TTFT 0.36s · measured
LFM2.5-24B-A2B Q4_K_M KV q8_0 · 3941 ctx	204	LFM2.5-24B-A2B Q4_K_M KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3941 prompt tokens · 4k-cached · total 1.3 · TTFT 0.02s · measured
LFM2.5-24B-A2B Q4_K_M KV q8_0 · 15637 ctx	188	LFM2.5-24B-A2B Q4_K_M KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cold · total 2.8 · TTFT 1.4s · measured
LFM2.5-24B-A2B Q4_K_M KV q8_0 · 15637 ctx	182.1	LFM2.5-24B-A2B Q4_K_M KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cached · total 1.5 · TTFT 0.05s · measured
LFM2.5-24B-A2B Q4_K_M KV q8_0 · 29425 ctx	185.5	LFM2.5-24B-A2B Q4_K_M KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29425 prompt tokens · 30k-cold · total 4.1 · TTFT 2.7s · measured
LFM2.5-24B-A2B Q4_K_M KV q8_0 · 29425 ctx	178.2	LFM2.5-24B-A2B Q4_K_M KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29425 prompt tokens · 30k-cached · total 1.5 · TTFT 0.07s · measured
Qwen3.8-27B UD-Q4_K_S KV f16 · 81 ctx	31.4	Qwen3.8-27B UD-Q4_K_S KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 81 prompt tokens · fixed-0 · total 8.5 · TTFT 0.33s · measured
Qwen3.8-27B UD-Q4_K_S KV f16 · 81 ctx	31.4	Qwen3.8-27B UD-Q4_K_S KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 81 prompt tokens · fixed-1 · total 8.6 · TTFT 0.52s · measured
Qwen3.8-27B UD-Q4_K_S KV f16 · 81 ctx	31.5	Qwen3.8-27B UD-Q4_K_S KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 81 prompt tokens · fixed-2 · total 8.9 · TTFT 0.78s · measured
Qwen3.8-27B UD-Q4_K_S KV f16 · 3927 ctx	30.7	Qwen3.8-27B UD-Q4_K_S KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 3927 prompt tokens · 4k-cold · total 12 · TTFT 3.72s · measured
Qwen3.8-27B UD-Q4_K_S KV f16 · 3927 ctx	30.8	Qwen3.8-27B UD-Q4_K_S KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 3927 prompt tokens · 4k-cached · total 8.6 · TTFT 0.34s · measured
Qwen3.8-27B UD-Q4_K_S KV f16 · 15624 ctx	29.7	Qwen3.8-27B UD-Q4_K_S KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 15624 prompt tokens · 16k-cold · total 22.4 · TTFT 13.84s · measured
Qwen3.8-27B UD-Q4_K_S KV f16 · 15624 ctx	29.7	Qwen3.8-27B UD-Q4_K_S KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 15624 prompt tokens · 16k-cached · total 9 · TTFT 0.38s · measured
Qwen3.8-27B UD-Q4_K_S KV f16 · 29408 ctx	28.6	Qwen3.8-27B UD-Q4_K_S KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 29408 prompt tokens · 30k-cold · total 36.8 · TTFT 27.9s · measured
Qwen3.8-27B UD-Q4_K_S KV f16 · 29408 ctx	28.3	Qwen3.8-27B UD-Q4_K_S KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 29408 prompt tokens · 30k-cached · total 9.5 · TTFT 0.51s · measured
Qwen3.6-27B UD-Q4_K_XL KV f16 · 39 ctx	28.7	Qwen3.6-27B UD-Q4_K_XL KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 39 prompt tokens · fixed-0 · total 9.3 · TTFT 0.42s · measured
Qwen3.6-27B UD-Q4_K_XL KV f16 · 39 ctx	28.6	Qwen3.6-27B UD-Q4_K_XL KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 39 prompt tokens · fixed-1 · total 9.3 · TTFT 0.43s · measured
Qwen3.6-27B UD-Q4_K_XL KV f16 · 39 ctx	28.5	Qwen3.6-27B UD-Q4_K_XL KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 39 prompt tokens · fixed-2 · total 9.2 · TTFT 0.22s · measured
Qwen3.6-27B UD-Q4_K_XL KV f16 · 3937 ctx	28.2	Qwen3.6-27B UD-Q4_K_XL KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 3937 prompt tokens · 4k-cold · total 13.1 · TTFT 4.02s · measured
Qwen3.6-27B UD-Q4_K_XL KV f16 · 3937 ctx	28.3	Qwen3.6-27B UD-Q4_K_XL KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 3937 prompt tokens · 4k-cached · total 9.3 · TTFT 0.26s · measured
Qwen3.6-27B UD-Q4_K_XL KV f16 · 15634 ctx	27.1	Qwen3.6-27B UD-Q4_K_XL KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 15634 prompt tokens · 16k-cold · total 24.7 · TTFT 15.25s · measured
Qwen3.6-27B UD-Q4_K_XL KV f16 · 15634 ctx	27.2	Qwen3.6-27B UD-Q4_K_XL KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 15634 prompt tokens · 16k-cached · total 9.9 · TTFT 0.52s · measured
Qwen3.6-27B UD-Q4_K_XL KV f16 · 29420 ctx	26.3	Qwen3.6-27B UD-Q4_K_XL KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 29420 prompt tokens · 30k-cold · total 39.7 · TTFT 29.99s · measured
Qwen3.6-27B UD-Q4_K_XL KV f16 · 29420 ctx	26.2	Qwen3.6-27B UD-Q4_K_XL KV f16 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 29420 prompt tokens · 30k-cached · total 10.2 · TTFT 0.42s · measured
LFM2.5-24B-A2B Q4_K_M KV q4_0 · 39 ctx	199.2	LFM2.5-24B-A2B Q4_K_M KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 39 prompt tokens · fixed-0 · total 1.6 · TTFT 0.35s · measured
LFM2.5-24B-A2B Q4_K_M KV q4_0 · 39 ctx	203.8	LFM2.5-24B-A2B Q4_K_M KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 39 prompt tokens · fixed-1 · total 1.3 · TTFT 0.03s · measured
LFM2.5-24B-A2B Q4_K_M KV q4_0 · 39 ctx	161	LFM2.5-24B-A2B Q4_K_M KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 39 prompt tokens · fixed-2 · total 1.6 · TTFT 0.02s · measured
LFM2.5-24B-A2B Q4_K_M KV q4_0 · 3941 ctx	182.6	LFM2.5-24B-A2B Q4_K_M KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3941 prompt tokens · 4k-cold · total 1.7 · TTFT 0.35s · measured
LFM2.5-24B-A2B Q4_K_M KV q4_0 · 3941 ctx	214	LFM2.5-24B-A2B Q4_K_M KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3941 prompt tokens · 4k-cached · total 1.2 · TTFT 0.02s · measured
LFM2.5-24B-A2B Q4_K_M KV q4_0 · 15637 ctx	186.7	LFM2.5-24B-A2B Q4_K_M KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cold · total 2.7 · TTFT 1.33s · measured
LFM2.5-24B-A2B Q4_K_M KV q4_0 · 15637 ctx	191.3	LFM2.5-24B-A2B Q4_K_M KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cached · total 1.4 · TTFT 0.04s · measured
LFM2.5-24B-A2B Q4_K_M KV q4_0 · 29425 ctx	192.1	LFM2.5-24B-A2B Q4_K_M KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29425 prompt tokens · 30k-cold · total 4 · TTFT 2.68s · measured
LFM2.5-24B-A2B Q4_K_M KV q4_0 · 29425 ctx	187.7	LFM2.5-24B-A2B Q4_K_M KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29425 prompt tokens · 30k-cached · total 1.4 · TTFT 0.06s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · 38 ctx	366.8	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 38 prompt tokens · fixed-0 · total 1.1 · TTFT 0.4s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · 38 ctx	350.3	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 38 prompt tokens · fixed-1 · total 0.8 · TTFT 0.03s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · 38 ctx	371.2	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 38 prompt tokens · fixed-2 · total 0.7 · TTFT 0.02s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · 3940 ctx	329.8	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3940 prompt tokens · 4k-cold · total 1 · TTFT 0.22s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · 3940 ctx	367	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3940 prompt tokens · 4k-cached · total 0.7 · TTFT 0.02s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · 15637 ctx	315.6	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cold · total 1.6 · TTFT 0.75s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · 15637 ctx	360.5	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cached · total 0.8 · TTFT 0.04s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · 29426 ctx	309.6	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29426 prompt tokens · 30k-cold · total 2.3 · TTFT 1.5s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · 29426 ctx	329.1	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29426 prompt tokens · 30k-cached · total 0.8 · TTFT 0.05s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · 38 ctx	351.2	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 38 prompt tokens · fixed-0 · total 0.9 · TTFT 0.19s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · 38 ctx	355	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 38 prompt tokens · fixed-1 · total 0.7 · TTFT 0.02s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · 38 ctx	340.5	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 38 prompt tokens · fixed-2 · total 0.8 · TTFT 0.02s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · 3940 ctx	287.4	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3940 prompt tokens · 4k-cold · total 1.1 · TTFT 0.2s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · 3940 ctx	342.9	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3940 prompt tokens · 4k-cached · total 0.8 · TTFT 0.03s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · 15637 ctx	347.4	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cold · total 1.5 · TTFT 0.73s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · 15637 ctx	327.3	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cached · total 0.8 · TTFT 0.04s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · 29426 ctx	315.2	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29426 prompt tokens · 30k-cold · total 2.2 · TTFT 1.39s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · 29426 ctx	293.4	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29426 prompt tokens · 30k-cached · total 0.9 · TTFT 0.06s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · 38 ctx	374.4	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 38 prompt tokens · fixed-0 · total 0.9 · TTFT 0.25s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · 38 ctx	382.7	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 38 prompt tokens · fixed-1 · total 0.7 · TTFT 0.02s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · 38 ctx	386.9	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 38 prompt tokens · fixed-2 · total 0.7 · TTFT 0.02s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · 3940 ctx	352.2	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3940 prompt tokens · 4k-cold · total 0.9 · TTFT 0.19s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · 3940 ctx	369.3	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3940 prompt tokens · 4k-cached · total 0.7 · TTFT 0.02s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · 15637 ctx	365.9	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cold · total 1.4 · TTFT 0.67s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · 15637 ctx	334.8	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15637 prompt tokens · 16k-cached · total 0.8 · TTFT 0.04s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · 29426 ctx	321.7	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29426 prompt tokens · 30k-cold · total 2.1 · TTFT 1.35s · measured
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · 29426 ctx	317.3	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29426 prompt tokens · 30k-cached · total 0.9 · TTFT 0.05s · measured
Gemma-4-12b-it UD-Q4_K_XL KV f16 · 45 ctx	90.8	Gemma-4-12b-it UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 45 prompt tokens · fixed-0 · total 3 · TTFT 0.14s · measured
Gemma-4-12b-it UD-Q4_K_XL KV f16 · 45 ctx	90.8	Gemma-4-12b-it UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 45 prompt tokens · fixed-1 · total 3.1 · TTFT 0.25s · measured
Gemma-4-12b-it UD-Q4_K_XL KV f16 · 45 ctx	91	Gemma-4-12b-it UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 45 prompt tokens · fixed-2 · total 3 · TTFT 0.19s · measured
Gemma-4-12b-it UD-Q4_K_XL KV f16 · 3936 ctx	87.8	Gemma-4-12b-it UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3936 prompt tokens · 4k-cold · total 4.8 · TTFT 1.91s · measured
Gemma-4-12b-it UD-Q4_K_XL KV f16 · 3936 ctx	86.4	Gemma-4-12b-it UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 3936 prompt tokens · 4k-cached · total 3.3 · TTFT 0.38s · measured
Gemma-4-12b-it UD-Q4_K_XL KV f16 · 15632 ctx	85.1	Gemma-4-12b-it UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15632 prompt tokens · 16k-cold · total 6.9 · TTFT 3.92s · measured
Gemma-4-12b-it UD-Q4_K_XL KV f16 · 15632 ctx	84	Gemma-4-12b-it UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 15632 prompt tokens · 16k-cached · total 3.6 · TTFT 0.52s · measured
Gemma-4-12b-it UD-Q4_K_XL KV f16 · 29417 ctx	84.7	Gemma-4-12b-it UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29417 prompt tokens · 30k-cold · total 9.8 · TTFT 6.83s · measured
Gemma-4-12b-it UD-Q4_K_XL KV f16 · 29417 ctx	83.8	Gemma-4-12b-it UD-Q4_K_XL KV f16 · RTX 4090 24 GiB · llama.cpp-cuda · 1 cards (RTX 4090 24 GiB · llama.cpp-cuda · 1 card) · 29417 prompt tokens · 30k-cached · total 3.5 · TTFT 0.5s · measured
Qwen3.6-27B UD-Q4_K_XL KV q8_0 · 39 ctx	28	Qwen3.6-27B UD-Q4_K_XL KV q8_0 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 39 prompt tokens · fixed-0 · total 9.5 · TTFT 0.41s · measured
Qwen3.6-27B UD-Q4_K_XL KV q8_0 · 39 ctx	27.8	Qwen3.6-27B UD-Q4_K_XL KV q8_0 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 39 prompt tokens · fixed-1 · total 9.6 · TTFT 0.43s · measured
Qwen3.6-27B UD-Q4_K_XL KV q8_0 · 39 ctx	27.8	Qwen3.6-27B UD-Q4_K_XL KV q8_0 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 39 prompt tokens · fixed-2 · total 9.4 · TTFT 0.28s · measured
Qwen3.6-27B UD-Q4_K_XL KV q8_0 · 3937 ctx	27.2	Qwen3.6-27B UD-Q4_K_XL KV q8_0 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 3937 prompt tokens · 4k-cold · total 13.5 · TTFT 4.14s · measured
Qwen3.6-27B UD-Q4_K_XL KV q8_0 · 3937 ctx	27.3	Qwen3.6-27B UD-Q4_K_XL KV q8_0 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 3937 prompt tokens · 4k-cached · total 9.7 · TTFT 0.41s · measured
Qwen3.6-27B UD-Q4_K_XL KV q8_0 · 15634 ctx	25.2	Qwen3.6-27B UD-Q4_K_XL KV q8_0 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 15634 prompt tokens · 16k-cold · total 25.2 · TTFT 15.1s · measured
Qwen3.6-27B UD-Q4_K_XL KV q8_0 · 15634 ctx	25.2	Qwen3.6-27B UD-Q4_K_XL KV q8_0 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 15634 prompt tokens · 16k-cached · total 10.7 · TTFT 0.55s · measured
Qwen3.6-27B UD-Q4_K_XL KV q8_0 · 29420 ctx	23.4	Qwen3.6-27B UD-Q4_K_XL KV q8_0 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 29420 prompt tokens · 30k-cold · total 40.8 · TTFT 29.88s · measured
Qwen3.6-27B UD-Q4_K_XL KV q8_0 · 29420 ctx	23.4	Qwen3.6-27B UD-Q4_K_XL KV q8_0 · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 cards (AMD R9700 32 GiB (1 card) · llama.cpp-rocm · 1 card) · 29420 prompt tokens · 30k-cached · total 11.7 · TTFT 0.74s · measured

Suite score and published 95% interval

score (0–1) · first 24 of 73 rows; all values in the table

Chart data and comparison dimensions

Benchmark family coverage

Recorded sample counts by model and family; a recorded sample does not establish qualification. Exact pass counts are in the table.

Coverage data

Benchmark family coverage — recorded observations
Configuration	Suite / family	Samples
Claude Fable 5.1	coding	14
Claude Fable 5.1	logic	16
Claude Fable 5.1	review	21
Claude Fable 5.1	science	48
Claude Fable 5.1	sql	30
Claude Opus 5.5	coding	14
Claude Opus 5.5	logic	16
Claude Opus 5.5	review	21
Claude Opus 5.5	science	48
Claude Opus 5.5	sql	30
Claude Sonnet 5.5	coding	14
Claude Sonnet 5.5	logic	16
Claude Sonnet 5.5	review	21
Claude Sonnet 5.5	science	48
Claude Sonnet 5.5	sql	30
GPT 5.6 Terra	coding	14
GPT 5.6 Terra	logic	16
GPT 5.6 Terra	review	21
GPT 5.6 Terra	science	48
GPT 5.6 Terra	sql	30
GPT 6 Astra	coding	14
GPT 6 Astra	logic	16
GPT 6 Astra	review	21
GPT 6 Astra	science	48
GPT 6 Astra	sql	30
GPT 6 Luna	coding	14
GPT 6 Luna	logic	16
GPT 6 Luna	review	21
GPT 6 Luna	science	48
GPT 6 Luna	sql	30
GPT 6.1 Sol	coding	14
GPT 6.1 Sol	logic	16
GPT 6.1 Sol	review	21
GPT 6.1 Sol	science	48
GPT 6.1 Sol	sql	30
Gemma-4-12b-it UD-Q4_K_XL KV f16	coding	14
Gemma-4-12b-it UD-Q4_K_XL KV f16	logic	16
Gemma-4-12b-it UD-Q4_K_XL KV f16	mcp	15
Gemma-4-12b-it UD-Q4_K_XL KV f16	review	42
Gemma-4-12b-it UD-Q4_K_XL KV f16	science	48
Gemma-4-12b-it UD-Q4_K_XL KV f16	sql	60
Gemma-4-12b-it UD-Q4_K_XL KV q4_0	review	7
Gemma-4-12b-it UD-Q4_K_XL KV q4_0	sql	10
Gemma-4-12b-it UD-Q4_K_XL KV q8_0	review	12
Gemma-4-12b-it UD-Q4_K_XL KV q8_0	sql	10
Gemma-4-31B-it UD-Q4_K_XL KV f16	review	28
Gemma-4-31B-it UD-Q4_K_XL KV f16	sql	60
LFM2.5-24B-A2B Q4_K_M KV f16	coding	14
LFM2.5-24B-A2B Q4_K_M KV f16	logic	16
LFM2.5-24B-A2B Q4_K_M KV f16	mcp	15
LFM2.5-24B-A2B Q4_K_M KV f16	review	42
LFM2.5-24B-A2B Q4_K_M KV f16	science	48
LFM2.5-24B-A2B Q4_K_M KV f16	sql	60
LFM2.5-24B-A2B Q4_K_M KV q4_0	coding	14
LFM2.5-24B-A2B Q4_K_M KV q4_0	logic	16
LFM2.5-24B-A2B Q4_K_M KV q4_0	mcp	15
LFM2.5-24B-A2B Q4_K_M KV q4_0	review	7
LFM2.5-24B-A2B Q4_K_M KV q4_0	science	48
LFM2.5-24B-A2B Q4_K_M KV q4_0	sql	10
LFM2.5-24B-A2B Q4_K_M KV q8_0	coding	14
LFM2.5-24B-A2B Q4_K_M KV q8_0	logic	16
LFM2.5-24B-A2B Q4_K_M KV q8_0	mcp	15
LFM2.5-24B-A2B Q4_K_M KV q8_0	review	7
LFM2.5-24B-A2B Q4_K_M KV q8_0	science	48
LFM2.5-24B-A2B Q4_K_M KV q8_0	sql	10
LFM2.5-8B-A1B UD-Q4_K_XL KV f16	coding	14
LFM2.5-8B-A1B UD-Q4_K_XL KV f16	logic	16
LFM2.5-8B-A1B UD-Q4_K_XL KV f16	mcp	15
LFM2.5-8B-A1B UD-Q4_K_XL KV f16	review	42
LFM2.5-8B-A1B UD-Q4_K_XL KV f16	science	48
LFM2.5-8B-A1B UD-Q4_K_XL KV f16	sql	60
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0	coding	14
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0	logic	16
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0	mcp	15
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0	review	7
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0	science	48
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0	sql	10
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0	coding	14
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0	logic	16
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0	mcp	15
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0	review	7
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0	science	48
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0	sql	10
Qwen3.5-27B UD-Q4_K_XL KV f16	review	36
Qwen3.5-27B UD-Q4_K_XL KV f16	sql	60
Qwen3.6-27B UD-Q4_K_XL KV f16	coding	14
Qwen3.6-27B UD-Q4_K_XL KV f16	lab	12
Qwen3.6-27B UD-Q4_K_XL KV f16	logic	16
Qwen3.6-27B UD-Q4_K_XL KV f16	mcp	15
Qwen3.6-27B UD-Q4_K_XL KV f16	review	42
Qwen3.6-27B UD-Q4_K_XL KV f16	science	48
Qwen3.6-27B UD-Q4_K_XL KV f16	sql	60
Qwen3.6-27B UD-Q4_K_XL KV q4_0	review	7
Qwen3.6-27B UD-Q4_K_XL KV q4_0	sql	10
Qwen3.6-27B UD-Q4_K_XL KV q8_0	review	7
Qwen3.6-27B UD-Q4_K_XL KV q8_0	sql	10
Qwen3.8-27B UD-Q4_K_S KV f16	coding	14
Qwen3.8-27B UD-Q4_K_S KV f16	lab	12
Qwen3.8-27B UD-Q4_K_S KV f16	logic	16
Qwen3.8-27B UD-Q4_K_S KV f16	mcp	15
Qwen3.8-27B UD-Q4_K_S KV f16	review	14
Qwen3.8-27B UD-Q4_K_S KV f16	science	48
Qwen3.8-27B UD-Q4_K_S KV f16	sql	20
Qwen3.8-27B UD-Q4_K_XL KV f16	review	38
Qwen3.8-27B UD-Q4_K_XL KV f16	sql	60

Published lane progress

completed units

Chart data and comparison dimensions

Published lane progress — exact data
Group	Value (completed units)	Dimensions and source
Lane 1	87	87 done · 0 pending · 0 retrying · 87 units
Lane 2	155	155 done · 211 pending · 0 retrying · 366 units
Lane 3	303	303 done · 0 pending · 1 retrying · 304 units
Lane 4	140	140 done · 0 pending · 7 retrying · 147 units

Observed model placements

cards

Chart data and comparison dimensions

Observed model placements — exact data
Group	Value (cards)	Dimensions and source
Gemma-4-12b-it	1	UD-Q4_K_XL · RTX 4090 24 GiB · llama.cpp-cuda · measured · 2026-10-25T23:03:58.394274969Z
Gemma-4-31B-it	1	UD-Q4_K_XL · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · measured · 2026-10-25T01:09:24.512369098Z
LFM2.5-24B-A2B	1	Q4_K_M · RTX 4090 24 GiB · llama.cpp-cuda · measured · 2026-10-25T05:33:37.886905249Z
LFM2.5-8B-A1B	1	UD-Q4_K_XL · RTX 4090 24 GiB · llama.cpp-cuda · measured · 2026-10-25T07:54:48.294699165Z
Qwen3.5-27B	1	UD-Q4_K_XL · RTX 4090 24 GiB · llama.cpp-cuda · measured · 2026-10-26T06:50:41.253546576Z
Qwen3.6-27B	1	UD-Q4_K_XL · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · measured · 2026-10-25T20:59:17.363160309Z
Qwen3.8-27B	1	UD-Q4_K_XL · AMD R9700 32 GiB (1 card) · llama.cpp-rocm · measured · 2026-10-25T22:51:54.660018136Z

Active lanes

wired Content-free export; private host labels omitted. Exported speed groups do not establish exact fingerprint/context comparability or qualification.

Lane	Done	Pending	Retrying	Units
Lane 1	87	0	0	87
Lane 2	155	211	0	366
Lane 3	303	0	1	304
Lane 4	140	0	7	147

Recent observations

wired Content-free export; private host labels omitted. Exported speed groups do not establish exact fingerprint/context comparability or qualification.

Latest observed placements; this snapshot does not contain a result-change history.

Model	Quantization	Hardware	Runtime	Cards	Recorded at
Qwen3.5-27B	UD-Q4_K_XL	RTX 4090 24 GiB	llama.cpp-cuda	1	2026-10-26T06:50:41.253546576Z
Gemma-4-12b-it	UD-Q4_K_XL	RTX 4090 24 GiB	llama.cpp-cuda	1	2026-10-25T23:03:58.394274969Z
Qwen3.8-27B	UD-Q4_K_XL	AMD R9700 32 GiB (1 card)	llama.cpp-rocm	1	2026-10-25T22:51:54.660018136Z
Qwen3.6-27B	UD-Q4_K_XL	AMD R9700 32 GiB (1 card)	llama.cpp-rocm	1	2026-10-25T20:59:17.363160309Z
LFM2.5-8B-A1B	UD-Q4_K_XL	RTX 4090 24 GiB	llama.cpp-cuda	1	2026-10-25T07:54:48.294699165Z
LFM2.5-24B-A2B	Q4_K_M	RTX 4090 24 GiB	llama.cpp-cuda	1	2026-10-25T05:33:37.886905249Z
Gemma-4-31B-it	UD-Q4_K_XL	AMD R9700 32 GiB (1 card)	llama.cpp-rocm	1	2026-10-25T01:09:24.512369098Z

Group	Value (score (0–1))	Low	High	Dimensions and source
Claude Sonnet 5.5 · overall	0.96	0.92	0.99	Claude Sonnet 5.5 · overall · n=78 · 550 tokens/sample · published 95% interval; calculation method not exported
Claude Sonnet 5.5 · coding	1	1	1	Claude Sonnet 5.5 · coding · n=14 · 512 tokens/sample · published 95% interval; calculation method not exported
Claude Sonnet 5.5 · logic	1	1	1	Claude Sonnet 5.5 · logic · n=16 · 1186 tokens/sample · published 95% interval; calculation method not exported
Claude Sonnet 5.5 · science	0.88	0.77	0.96	Claude Sonnet 5.5 · science · n=48 · 350 tokens/sample · published 95% interval; calculation method not exported
Claude Opus 5.5 · overall	0.98	0.95	1	Claude Opus 5.5 · overall · n=78 · 615 tokens/sample · published 95% interval; calculation method not exported
Claude Opus 5.5 · coding	1	1	1	Claude Opus 5.5 · coding · n=14 · 590 tokens/sample · published 95% interval; calculation method not exported
Claude Opus 5.5 · logic	1	1	1	Claude Opus 5.5 · logic · n=16 · 1122 tokens/sample · published 95% interval; calculation method not exported
Claude Opus 5.5 · science	0.94	0.85	1	Claude Opus 5.5 · science · n=48 · 453 tokens/sample · published 95% interval; calculation method not exported
Claude Fable 5.1 · overall	0.97	0.93	0.99	Claude Fable 5.1 · overall · n=78 · 525 tokens/sample · published 95% interval; calculation method not exported
Claude Fable 5.1 · coding	1	1	1	Claude Fable 5.1 · coding · n=14 · 541 tokens/sample · published 95% interval; calculation method not exported
Claude Fable 5.1 · logic	1	1	1	Claude Fable 5.1 · logic · n=16 · 1065 tokens/sample · published 95% interval; calculation method not exported
Claude Fable 5.1 · science	0.9	0.79	0.98	Claude Fable 5.1 · science · n=48 · 340 tokens/sample · published 95% interval; calculation method not exported
GPT 6.1 Sol · overall	0.97	0.93	0.99	GPT 6.1 Sol · overall · n=78 · 323 tokens/sample · published 95% interval; calculation method not exported
GPT 6.1 Sol · coding	1	1	1	GPT 6.1 Sol · coding · n=14 · 333 tokens/sample · published 95% interval; calculation method not exported
GPT 6.1 Sol · logic	1	1	1	GPT 6.1 Sol · logic · n=16 · 647 tokens/sample · published 95% interval; calculation method not exported
GPT 6.1 Sol · science	0.9	0.79	0.98	GPT 6.1 Sol · science · n=48 · 213 tokens/sample · published 95% interval; calculation method not exported
GPT 6 Luna · overall	0.72	0.63	0.81	GPT 6 Luna · overall · n=78 · 257 tokens/sample · published 95% interval; calculation method not exported
GPT 6 Luna · coding	1	1	1	GPT 6 Luna · coding · n=14 · 298 tokens/sample · published 95% interval; calculation method not exported
GPT 6 Luna · logic	0.44	0.19	0.69	GPT 6 Luna · logic · n=16 · 620 tokens/sample · published 95% interval; calculation method not exported
GPT 6 Luna · science	0.71	0.56	0.83	GPT 6 Luna · science · n=48 · 124 tokens/sample · published 95% interval; calculation method not exported
GPT 6 Astra · overall	0.97	0.93	0.99	GPT 6 Astra · overall · n=78 · 282 tokens/sample · published 95% interval; calculation method not exported
GPT 6 Astra · coding	1	1	1	GPT 6 Astra · coding · n=14 · 308 tokens/sample · published 95% interval; calculation method not exported
GPT 6 Astra · logic	1	1	1	GPT 6 Astra · logic · n=16 · 552 tokens/sample · published 95% interval; calculation method not exported
GPT 6 Astra · science	0.9	0.79	0.98	GPT 6 Astra · science · n=48 · 184 tokens/sample · published 95% interval; calculation method not exported
GPT 5.6 Terra · overall	0.94	0.88	0.98	GPT 5.6 Terra · overall · n=78 · 349 tokens/sample · published 95% interval; calculation method not exported
GPT 5.6 Terra · coding	1	1	1	GPT 5.6 Terra · coding · n=14 · 366 tokens/sample · published 95% interval; calculation method not exported
GPT 5.6 Terra · logic	0.94	0.81	1	GPT 5.6 Terra · logic · n=16 · 547 tokens/sample · published 95% interval; calculation method not exported
GPT 5.6 Terra · science	0.88	0.77	0.96	GPT 5.6 Terra · science · n=48 · 279 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV f16 · overall	0.82	0.75	0.89	LFM2.5-24B-A2B Q4_K_M KV f16 · overall · n=186 · 7271 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV f16 · coding	0.82	0.61	1	LFM2.5-24B-A2B Q4_K_M KV f16 · coding · n=28 · 6718 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV f16 · logic	0.97	0.91	1	LFM2.5-24B-A2B Q4_K_M KV f16 · logic · n=32 · 7674 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV f16 · mcp	0.77	0.6	0.9	LFM2.5-24B-A2B Q4_K_M KV f16 · mcp · n=30 · 1910 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV f16 · science	0.73	0.61	0.84	LFM2.5-24B-A2B Q4_K_M KV f16 · science · n=96 · 8974 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV q8_0 · overall	0.87	0.79	0.93	LFM2.5-24B-A2B Q4_K_M KV q8_0 · overall · n=186 · 6504 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV q8_0 · coding	0.86	0.64	1	LFM2.5-24B-A2B Q4_K_M KV q8_0 · coding · n=28 · 6028 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV q8_0 · logic	0.97	0.91	1	LFM2.5-24B-A2B Q4_K_M KV q8_0 · logic · n=32 · 7932 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV q8_0 · mcp	0.87	0.7	1	LFM2.5-24B-A2B Q4_K_M KV q8_0 · mcp · n=30 · 1846 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV q8_0 · science	0.78	0.66	0.9	LFM2.5-24B-A2B Q4_K_M KV q8_0 · science · n=96 · 7623 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV q4_0 · overall	0.85	0.78	0.91	LFM2.5-24B-A2B Q4_K_M KV q4_0 · overall · n=186 · 6845 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV q4_0 · coding	0.86	0.68	1	LFM2.5-24B-A2B Q4_K_M KV q4_0 · coding · n=28 · 6097 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV q4_0 · logic	0.91	0.75	1	LFM2.5-24B-A2B Q4_K_M KV q4_0 · logic · n=32 · 8234 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV q4_0 · mcp	0.87	0.67	1	LFM2.5-24B-A2B Q4_K_M KV q4_0 · mcp · n=30 · 1627 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-24B-A2B Q4_K_M KV q4_0 · science	0.77	0.66	0.88	LFM2.5-24B-A2B Q4_K_M KV q4_0 · science · n=96 · 8231 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · overall	0.47	0.38	0.57	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · overall · n=186 · 5248 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · coding	0.54	0.29	0.75	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · coding · n=28 · 2780 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · logic	0.25	0.09	0.44	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · logic · n=32 · 14771 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · mcp	0.57	0.37	0.77	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · mcp · n=30 · 938 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · science	0.54	0.42	0.67	LFM2.5-8B-A1B UD-Q4_K_XL KV f16 · science · n=96 · 4140 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · overall	0.55	0.45	0.64	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · overall · n=186 · 5298 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · coding	0.64	0.39	0.86	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · coding · n=28 · 3168 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · logic	0.31	0.13	0.5	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · logic · n=32 · 15518 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · mcp	0.63	0.43	0.83	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · mcp · n=30 · 1564 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · science	0.59	0.47	0.72	LFM2.5-8B-A1B UD-Q4_K_XL KV q8_0 · science · n=96 · 3680 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · overall	0.54	0.45	0.62	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · overall · n=186 · 5800 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · coding	0.57	0.36	0.79	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · coding · n=28 · 2949 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · logic	0.28	0.16	0.44	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · logic · n=32 · 17785 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · mcp	0.63	0.4	0.83	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · mcp · n=30 · 853 tokens/sample · published 95% interval; calculation method not exported
LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · science	0.67	0.55	0.78	LFM2.5-8B-A1B UD-Q4_K_XL KV q4_0 · science · n=96 · 4183 tokens/sample · published 95% interval; calculation method not exported
Qwen3.8-27B UD-Q4_K_S KV f16 · overall	0.9	0.84	0.95	Qwen3.8-27B UD-Q4_K_S KV f16 · overall · n=186 · 4207 tokens/sample · published 95% interval; calculation method not exported
Qwen3.8-27B UD-Q4_K_S KV f16 · coding	0.82	0.61	1	Qwen3.8-27B UD-Q4_K_S KV f16 · coding · n=28 · 8748 tokens/sample · published 95% interval; calculation method not exported
Qwen3.8-27B UD-Q4_K_S KV f16 · logic	0.97	0.91	1	Qwen3.8-27B UD-Q4_K_S KV f16 · logic · n=32 · 2507 tokens/sample · published 95% interval; calculation method not exported
Qwen3.8-27B UD-Q4_K_S KV f16 · mcp	0.97	0.9	1	Qwen3.8-27B UD-Q4_K_S KV f16 · mcp · n=30 · 712 tokens/sample · published 95% interval; calculation method not exported
Qwen3.8-27B UD-Q4_K_S KV f16 · science	0.84	0.74	0.93	Qwen3.8-27B UD-Q4_K_S KV f16 · science · n=96 · 4541 tokens/sample · published 95% interval; calculation method not exported
Gemma-4-12b-it UD-Q4_K_XL KV f16 · overall	0.88	0.82	0.93	Gemma-4-12b-it UD-Q4_K_XL KV f16 · overall · n=186 · 8876 tokens/sample · published 95% interval; calculation method not exported
Gemma-4-12b-it UD-Q4_K_XL KV f16 · coding	0.82	0.61	1	Gemma-4-12b-it UD-Q4_K_XL KV f16 · coding · n=28 · 9956 tokens/sample · published 95% interval; calculation method not exported
Gemma-4-12b-it UD-Q4_K_XL KV f16 · logic	0.97	0.91	1	Gemma-4-12b-it UD-Q4_K_XL KV f16 · logic · n=32 · 9879 tokens/sample · published 95% interval; calculation method not exported
Gemma-4-12b-it UD-Q4_K_XL KV f16 · mcp	0.97	0.9	1	Gemma-4-12b-it UD-Q4_K_XL KV f16 · mcp · n=30 · 903 tokens/sample · published 95% interval; calculation method not exported
Gemma-4-12b-it UD-Q4_K_XL KV f16 · science	0.75	0.63	0.85	Gemma-4-12b-it UD-Q4_K_XL KV f16 · science · n=96 · 10719 tokens/sample · published 95% interval; calculation method not exported
Qwen3.6-27B UD-Q4_K_XL KV f16 · overall	0.91	0.85	0.96	Qwen3.6-27B UD-Q4_K_XL KV f16 · overall · n=186 · 7853 tokens/sample · published 95% interval; calculation method not exported
Qwen3.6-27B UD-Q4_K_XL KV f16 · coding	0.89	0.71	1	Qwen3.6-27B UD-Q4_K_XL KV f16 · coding · n=28 · 10898 tokens/sample · published 95% interval; calculation method not exported
Qwen3.6-27B UD-Q4_K_XL KV f16 · logic	0.94	0.81	1	Qwen3.6-27B UD-Q4_K_XL KV f16 · logic · n=32 · 11029 tokens/sample · published 95% interval; calculation method not exported
Qwen3.6-27B UD-Q4_K_XL KV f16 · mcp	1	1	1	Qwen3.6-27B UD-Q4_K_XL KV f16 · mcp · n=30 · 1086 tokens/sample · published 95% interval; calculation method not exported
Qwen3.6-27B UD-Q4_K_XL KV f16 · science	0.8	0.69	0.91	Qwen3.6-27B UD-Q4_K_XL KV f16 · science · n=96 · 8021 tokens/sample · published 95% interval; calculation method not exported