Shared results · Qwen3.8-Flash-Next quant comparison

Precision vs size: Qwen3.8-Flash-Next quants

Published 6 October 2026, updated 8 October 2026 as runs are recorded · 11 finished runs · 2 charts

hardware: AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory
protocol: KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.
runtime: 2 values; see the parameters of each run

proposed

Hypothesis

At equal GPU memory, agentionai's Gyro files stay closer to the original model (lower KL divergence, higher top-1 agreement) than other published quants; the Pareto frontier of fidelity vs GPU-resident size is made of Gyro and AP files.

Charts

KL divergence vs GPU memory (lower-left is better)

scatter · 11 runs plotted · 1 excluded

Colour: vendor. The points span 2 comparison contexts; each point's is in the source data.

Frontier chosen by hand: 3 runs, joined in GPU-resident weights (GiB) order. The dashed line connects recorded runs; it does not imply results between them.

How this chart is drawn

One point per successful run, no aggregation. Context series separate experiments, recorded comparison context, input references and environments; matching metadata does not establish experimental equivalence.

Displayed Y range: 0.1123 to 0.5043 · X range: 27.57 to 67.36

Schema b2b235ad-fe84-4c78-9970-147532f665f1 · 11 eligible runs · 1 excluded · Live as of 2026-10-08T09:29:39.894Z

Chart display (this view only)Show point labels (run titles)Fit X axis to visible valuesFit Y axis to visible values

Unchecked numeric axes include zero. Fitted axes may omit zero; bars always start at zero. Hiding a series rescales the graph, not the source table. Labels may overlap; full titles remain in tooltips and the table.

Source data and excluded runs

Exact values, source revisions and context series
Run	Experiment	Revision	GPU-resident weights (GiB)	Mean KL divergence vs Q8_0	Series	Context	Frontier
agentionai Gyro-S (TQ1_0): fidelity vs size	Shared experiment	1	27.64	0.435165	vendor: agentionai	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp (Vulkan)"	Yes
agentionai Gyro-M (TQ2_0): fidelity vs size	Shared experiment	1	34.98	0.305558	vendor: agentionai	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp (Vulkan)"	Yes
agentionai AP-Q4_K_XL: fidelity vs size	Shared experiment	1	67.36	0.112272	vendor: agentionai	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp (Vulkan)"	Yes
ISTA-DASLab GSQ-RCO IQ2_XS: fidelity vs size	Shared experiment	1	36.52	0.420224	vendor: ISTA-DASLab	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp (Vulkan)"
ISTA-DASLab GSQ-RCO Q2_0: fidelity vs size	Shared experiment	2	35.03	0.467805	vendor: ISTA-DASLab	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp (Vulkan)"
unsloth UD-Q2_K_XL: fidelity vs size	Shared experiment	1	46.62	0.334233	vendor: unsloth	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"
unsloth UD-IQ3_XXS: fidelity vs size	Shared experiment	1	49.5	0.264302	vendor: unsloth	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"
unsloth UD-Q3_K_XL: fidelity vs size	Shared experiment	1	56.97	0.18548	vendor: unsloth	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"
ISTA-DASLab GSQ-RCO IQ3_XXS: fidelity vs size	Shared experiment	1	43.8	0.310866	vendor: ISTA-DASLab	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"
ISTA-DASLab GSQ-RCO IQ3_S: fidelity vs size	Shared experiment	1	51.04	0.203961	vendor: ISTA-DASLab	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"
ISTA-DASLab GSQ-RCO Coder IQ1_M: fidelity vs size	Shared experiment	1	27.57	0.504256	vendor: ISTA-DASLab	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"

d1bbd7c5-6b44-4c18-899b-a8d730231ac2: Superseded by excluded-from-comparison

Top-1 agreement vs GPU memory (upper-left is better)

scatter · 11 runs plotted · 1 excluded

Colour: vendor. The points span 2 comparison contexts; each point's is in the source data.

Frontier chosen by hand: 3 runs, joined in GPU-resident weights (GiB) order. The dashed line connects recorded runs; it does not imply results between them.

How this chart is drawn

Displayed Y range: 72.96 to 85.08 · X range: 27.57 to 67.36

Schema b2b235ad-fe84-4c78-9970-147532f665f1 · 11 eligible runs · 1 excluded · Live as of 2026-10-08T09:29:39.902Z

Chart display (this view only)Show point labels (run titles)Fit X axis to visible valuesFit Y axis to visible values

Source data and excluded runs

Exact values, source revisions and context series
Run	Experiment	Revision	GPU-resident weights (GiB)	Top-1 agreement with Q8_0 (%)	Series	Context	Frontier
agentionai Gyro-S (TQ1_0): fidelity vs size	Shared experiment	1	27.64	74.05	vendor: agentionai	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp (Vulkan)"	Yes
agentionai Gyro-M (TQ2_0): fidelity vs size	Shared experiment	1	34.98	78.117	vendor: agentionai	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp (Vulkan)"	Yes
agentionai AP-Q4_K_XL: fidelity vs size	Shared experiment	1	67.36	85.078	vendor: agentionai	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp (Vulkan)"	Yes
ISTA-DASLab GSQ-RCO IQ2_XS: fidelity vs size	Shared experiment	1	36.52	74.22	vendor: ISTA-DASLab	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp (Vulkan)"
ISTA-DASLab GSQ-RCO Q2_0: fidelity vs size	Shared experiment	2	35.03	73.928	vendor: ISTA-DASLab	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp (Vulkan)"
unsloth UD-Q2_K_XL: fidelity vs size	Shared experiment	1	46.62	76.61	vendor: unsloth	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"
unsloth UD-IQ3_XXS: fidelity vs size	Shared experiment	1	49.5	79.389	vendor: unsloth	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"
unsloth UD-Q3_K_XL: fidelity vs size	Shared experiment	1	56.97	81.931	vendor: unsloth	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"
ISTA-DASLab GSQ-RCO IQ3_XXS: fidelity vs size	Shared experiment	1	43.8	77.682	vendor: ISTA-DASLab	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"
ISTA-DASLab GSQ-RCO IQ3_S: fidelity vs size	Shared experiment	1	51.04	81.315	vendor: ISTA-DASLab	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"
ISTA-DASLab GSQ-RCO Coder IQ1_M: fidelity vs size	Shared experiment	1	27.57	72.957	vendor: ISTA-DASLab	hardware="AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory" · protocol="KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions." · runtime="agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c"

d1bbd7c5-6b44-4c18-899b-a8d730231ac2: Superseded by excluded-from-comparison

Results

Finished runs and what they recorded.
Run	kld mean ↓ better	kld median ↓ better	top1 pct (%) ↑ better	gpu resident gib (GiB) ↓ better	file gib (GiB)	kld code ↓ better	kld reasoning ↓ better	mmlu pro (%) ↑ better	measured yyyymmdd
agentionai AP-Q4_K_XL: fidelity vs size	0.112272	0.032691	85.078	67.36	94.18	0.042271	0.083649	81.43	20,260,900
agentionai Gyro-M (TQ2_0): fidelity vs size	0.305558	0.08974	78.117	34.98	85.65	0.121349	0.242431	80	20,261,000
agentionai Gyro-S (TQ1_0): fidelity vs size	0.435165	0.140503	74.05	27.64	54.47	0.188147	0.386166	78.1	20,261,000
ISTA-DASLab GSQ-RCO Coder IQ1_M: fidelity vs size	0.504256	0.166775	72.957	27.57	54.4	0.207332	0.40311	Missing	20,261,000
ISTA-DASLab GSQ-RCO IQ2_XS: fidelity vs size	0.420224	0.139201	74.22	36.52	63.34	0.186178	0.396463	Missing	20,260,900
ISTA-DASLab GSQ-RCO IQ3_S: fidelity vs size	0.203961	0.057718	81.315	51.04	77.88	0.094699	0.196154	Missing	20,261,000
ISTA-DASLab GSQ-RCO IQ3_XXS: fidelity vs size	0.310866	0.090238	77.682	43.8	70.63	0.130686	0.292607	Missing	20,261,000
ISTA-DASLab GSQ-RCO Q2_0: fidelity vs size	0.467805	0.14989	73.928	35.03	61.86	0.209322	0.458675	80.95	20,261,000
unsloth UD-IQ3_XXS: fidelity vs size	0.264302	0.075732	79.389	49.5	76.33	0.10821	0.243885	Missing	20,261,000
unsloth UD-Q2_K_XL: fidelity vs size	0.334233	0.108277	76.61	46.62	73.45	0.139229	0.317638	Missing	20,261,000
unsloth UD-Q3_K_XL: fidelity vs size	0.18548	0.052315	81.931	56.97	83.81	0.076849	0.166222	Missing	20,261,000

Finished runs and what they recorded.

agentionai AP-Q4_K_XL: fidelity vs sizekld mean 0.112272

kld mean ↓: 0.112272
kld median ↓: 0.032691
top1 pct (%) ↑: 85.078
gpu resident gib (GiB) ↓: 67.36
file gib (GiB): 94.18
kld code ↓: 0.042271
kld reasoning ↓: 0.083649
mmlu pro (%) ↑: 81.43
measured yyyymmdd: 20,260,900

agentionai Gyro-M (TQ2_0): fidelity vs sizekld mean 0.305558

kld mean ↓: 0.305558
kld median ↓: 0.08974
top1 pct (%) ↑: 78.117
gpu resident gib (GiB) ↓: 34.98
file gib (GiB): 85.65
kld code ↓: 0.121349
kld reasoning ↓: 0.242431
mmlu pro (%) ↑: 80
measured yyyymmdd: 20,261,000

agentionai Gyro-S (TQ1_0): fidelity vs sizekld mean 0.435165

kld mean ↓: 0.435165
kld median ↓: 0.140503
top1 pct (%) ↑: 74.05
gpu resident gib (GiB) ↓: 27.64
file gib (GiB): 54.47
kld code ↓: 0.188147
kld reasoning ↓: 0.386166
mmlu pro (%) ↑: 78.1
measured yyyymmdd: 20,261,000

ISTA-DASLab GSQ-RCO Coder IQ1_M: fidelity vs sizekld mean 0.504256

kld mean ↓: 0.504256
kld median ↓: 0.166775
top1 pct (%) ↑: 72.957
gpu resident gib (GiB) ↓: 27.57
file gib (GiB): 54.4
kld code ↓: 0.207332
kld reasoning ↓: 0.40311
mmlu pro (%) ↑: Missing
measured yyyymmdd: 20,261,000

ISTA-DASLab GSQ-RCO IQ2_XS: fidelity vs sizekld mean 0.420224

kld mean ↓: 0.420224
kld median ↓: 0.139201
top1 pct (%) ↑: 74.22
gpu resident gib (GiB) ↓: 36.52
file gib (GiB): 63.34
kld code ↓: 0.186178
kld reasoning ↓: 0.396463
mmlu pro (%) ↑: Missing
measured yyyymmdd: 20,260,900

ISTA-DASLab GSQ-RCO IQ3_S: fidelity vs sizekld mean 0.203961

kld mean ↓: 0.203961
kld median ↓: 0.057718
top1 pct (%) ↑: 81.315
gpu resident gib (GiB) ↓: 51.04
file gib (GiB): 77.88
kld code ↓: 0.094699
kld reasoning ↓: 0.196154
mmlu pro (%) ↑: Missing
measured yyyymmdd: 20,261,000

ISTA-DASLab GSQ-RCO IQ3_XXS: fidelity vs sizekld mean 0.310866

kld mean ↓: 0.310866
kld median ↓: 0.090238
top1 pct (%) ↑: 77.682
gpu resident gib (GiB) ↓: 43.8
file gib (GiB): 70.63
kld code ↓: 0.130686
kld reasoning ↓: 0.292607
mmlu pro (%) ↑: Missing
measured yyyymmdd: 20,261,000

ISTA-DASLab GSQ-RCO Q2_0: fidelity vs sizekld mean 0.467805

kld mean ↓: 0.467805
kld median ↓: 0.14989
top1 pct (%) ↑: 73.928
gpu resident gib (GiB) ↓: 35.03
file gib (GiB): 61.86
kld code ↓: 0.209322
kld reasoning ↓: 0.458675
mmlu pro (%) ↑: 80.95
measured yyyymmdd: 20,261,000

unsloth UD-IQ3_XXS: fidelity vs sizekld mean 0.264302

kld mean ↓: 0.264302
kld median ↓: 0.075732
top1 pct (%) ↑: 79.389
gpu resident gib (GiB) ↓: 49.5
file gib (GiB): 76.33
kld code ↓: 0.10821
kld reasoning ↓: 0.243885
mmlu pro (%) ↑: Missing
measured yyyymmdd: 20,261,000

unsloth UD-Q2_K_XL: fidelity vs sizekld mean 0.334233

kld mean ↓: 0.334233
kld median ↓: 0.108277
top1 pct (%) ↑: 76.61
gpu resident gib (GiB) ↓: 46.62
file gib (GiB): 73.45
kld code ↓: 0.139229
kld reasoning ↓: 0.317638
mmlu pro (%) ↑: Missing
measured yyyymmdd: 20,261,000

unsloth UD-Q3_K_XL: fidelity vs sizekld mean 0.18548

kld mean ↓: 0.18548
kld median ↓: 0.052315
top1 pct (%) ↑: 81.931
gpu resident gib (GiB) ↓: 56.97
file gib (GiB): 83.81
kld code ↓: 0.076849
kld reasoning ↓: 0.166222
mmlu pro (%) ↑: Missing
measured yyyymmdd: 20,261,000

The parameters of each run

Parameters each run was produced with, and the context it was measured in. Values marked "added later" were filled in after the run executed.
Run	experts kept	file	ours	vendor	hardware (context)	protocol (context)	runtime (context)
agentionai Gyro-S (TQ1_0): fidelity vs size	512	Gyro-S (TQ1_0)	true	agentionai	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp (Vulkan)
agentionai Gyro-M (TQ2_0): fidelity vs size	512	Gyro-M (TQ2_0)	true	agentionai	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp (Vulkan)
agentionai AP-Q4_K_XL: fidelity vs size	512	AP-Q4_K_XL	true	agentionai	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp (Vulkan)
ISTA-DASLab GSQ-RCO IQ2_XS: fidelity vs size	512	GSQ-RCO IQ2_XS	false	ISTA-DASLab	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp (Vulkan)
ISTA-DASLab GSQ-RCO Q2_0: fidelity vs size	512	GSQ-RCO Q2_0	false	ISTA-DASLab	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp (Vulkan)
unsloth UD-Q2_K_XL: fidelity vs size	512	UD-Q2_K_XL	false	unsloth	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c
unsloth UD-IQ3_XXS: fidelity vs size	512	UD-IQ3_XXS	false	unsloth	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c
unsloth UD-Q3_K_XL: fidelity vs size	512	UD-Q3_K_XL	false	unsloth	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c
ISTA-DASLab GSQ-RCO IQ3_XXS: fidelity vs size	512	GSQ-RCO IQ3_XXS	false	ISTA-DASLab	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c
ISTA-DASLab GSQ-RCO IQ3_S: fidelity vs size	512	GSQ-RCO IQ3_S	false	ISTA-DASLab	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c
ISTA-DASLab GSQ-RCO Coder IQ1_M: fidelity vs size	256	GSQ-RCO Coder IQ1_M	false	ISTA-DASLab	AMD Strix Halo (Radeon 8060S, gfx1151), 128 GB unified memory	KL divergence and top-1 token agreement against the unsloth Q8_0 original on 60 x 2,048 held-out tokens (2026 technical writing and code), llama-perplexity --kl-divergence, n-gram table read from disk. Code/reasoning: same measure on pagoda-task code and reasoning traces written by other models. MMLU-Pro: 70 questions x 3 repetitions.	agentionai/llama.cpp dev build (Vulkan), b11330 d8841825c

1 run is not shown: unfinished or replaced

DJLougen Mooney: fidelity vs size succeededsuperseded

How this was measured

Goal: Shareable comparison of published Qwen3.8-Flash-Next quantisations (agentionai and other vendors): output fidelity against the Q8_0 original versus GPU memory, behaviour, and speed by hardware and engine. Measurements and evaluation protocol only.