Head-to-head · ggufbench model-vs-model compare

Generated 14/10/2026 10:43 · scores from the harness functions (same math as the CLI) · single sample per task (n=1, fixed seed/temp): this page orders, it does not certify close calls
weights tools 0.3 · code 0.3 · agent 0.25 · speed 0.15 better green / worse red · ties stay neutral
pick a pair

strata-iq3_xxs think:off strata 0.1.43.1

0.80 ±2
global · rank #10 of 30 · measured 2026-10-13 · set 12/60/28
model_config_flags
qwen3.8-flash-next-iq3_xxs --pack ~/jordi/ia/llama/Strata-data/packs/iq3_xxs --native ~/jordi/ia/llama/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf --ple-gguf ~/jordi/ia/llama/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf --expert-profile ~/jordi/ia/llama/Strata/data/expert-profile.bin --expert-cache auto --vram-reserve-mib 1000 --prefill auto --spec 4 --mtp ~/jordi/ia/llama/Strata-data/mtp/rt --max-context 131072 --kv int8 --control-vector-scaled ~/jordi/ia/llama/Strata/data/experimental-speed-projection/Qwen3.8-Flash-Next-experimental-speed-projection.gguf:1.0 --control-vector-layer-range 4 44 --cvec-mode project --cvec-dir per-layer --spec-min-p 0.5 --pool-workers 4
environment
RTX 5080 16GB × 2 · Linux · 60 GB RAM

strata-iq3_xxs think:on strata 0.1.43.1

0.76 ±0
global · rank #23 of 30 · measured 2026-10-13 · set 12/60/28
model_config_flags
qwen3.8-flash-next-iq3_xxs --pack ~/jordi/ia/llama/Strata-data/packs/iq3_xxs --native ~/jordi/ia/llama/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf --ple-gguf ~/jordi/ia/llama/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf --expert-profile ~/jordi/ia/llama/Strata/data/expert-profile.bin --expert-cache auto --vram-reserve-mib 1000 --prefill auto --spec 4 --mtp ~/jordi/ia/llama/Strata-data/mtp/rt --max-context 131072 --kv int8 --control-vector-scaled ~/jordi/ia/llama/Strata/data/experimental-speed-projection/Qwen3.8-Flash-Next-experimental-speed-projection.gguf:1.0 --control-vector-layer-range 4 44 --cvec-mode project --cvec-dir per-layer --spec-min-p 0.5 --pool-workers 4
environment
RTX 5080 16GB × 2 · Linux · 60 GB RAM

Global +48%

Weighted globaltools .30 · code .30 · agent .25 · speed .15
0.80
0.76
Higher weighted global. +5.4%
Pool rankposition in the accumulated scoreboard; lower wins
#10
#23
Higher pool rank. +57%
Wall clock (total run)time to complete the whole benchmark in this mode (C8); lower wins
16m 48s
95m 51s
Less wall clock. +82%

Speed +53%

Prompt + reply latencyspeed-task round trip; lower is better
50 ms
610 ms
Faster prompt + reply latency. +92%
Generation throughputoutput tokens/s
99 t/s
101 t/s
Faster generation. +1.7%
Prompt processinginput tokens/s
140 t/s
121 t/s
Faster prompt processing. +15%

Tools +3.3%

Tool-calling score8 graded tasks: arguments, multi-call, error handling
86%
83%
Better tool-calling. +3.3%

Code +7.6%

Code overallall levels; (N excl.) = excluded with a recorded reason
88%
89%
Better code overall. +1.3%
Easy tierbase tasks
92%
88%
Better on easy code. +4.5%
Hard tier_hard tasks
94%
96%
Better on hard code. +2.1%
Expert tier_expert tasks (multi-file)
72%
86%
Better on expert code. +19%

Agent

Agent scoreturn + tool-call accounting over the agent tasks
96%
96%
tie