2026-10-28 · Windows · Ryzen 9 7950X3D · Radeon AI PRO R9700
A model scorecard round on AMD Ryzen 9 7950X3D 16-Core Processor · AMD Radeon AI PRO R9700 · 64 GB: 2 models across 4 task sets. Every round is listed in the model scorecard, with its results side by side in one table.
AMD Ryzen 9 7950X3D 16-Core Processor · AMD Radeon AI PRO R9700 · 64 GB · llama-cpp · 3 trials per task · 2026-10-28 · gezel 22fec99111a5 · catalog 0.3.1
258 frozen trials across four suites, three repetitions per model/scenario. Native 0.1.54, llama.cpp Vulkan, generalist auto. Six raw infra exclusions remain: three Qwen custom-gate Windows denyNet failures and three model-authored gate-configuration failures (two Gemma, one Qwen); the latter are not Windows failures. Qwen complex-work has no eligible custom-gate cell, so its suite headline is withheld. Baseline DocBlocks lacked its tools in all six trials. The separately labeled activation subset does not replace these outcomes. Task measurements include every attempted trial and use engine-log throughput, not admission probes; system RSS and GPU memory are separate. Review means use declared Codex inline axes or game-HTML Claude reviews, never both for the same trial. Six misaligned A/B reviews, seven incidental HTML reviews and 137 tasks with no applicable artifact review are excluded from these means. Both complete judge ledgers, corrected spreadsheet selections, capability-rubric differences, performance gaps and evaluator/setup findings are preserved in the R9700 downloadable report.
System memory counts this computer's own memory only. What the model keeps on the separate graphics card isn't included.
General capability
| Model | Size | Tasks passed | Context | System memory |
|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) (kv: f16) | 27B | 32/33 (97%) | 139K | 18.8 GB |
| Gemma 4 (12B, Q4) (kv: f16) | 12B | 27/33 (82%) | 256K | 16.3 GB |
Task performance
Task measurements across all attempted trials, including failures. Measurement methods and limits are stated in period labels and round notes. Values are per-trial medians (minimum–maximum; sample count), not admission probes. Wall time includes tools. System RSS and GPU memory are separate peak measurements; do not add their medians. Periods remain separate because hardware performance can change.
| Model | Period | Trials | Output (tok/s) | Prefill (tok/s) | Wall time (min) | Peak system RSS (GiB) | Peak GPU memory (GiB) |
|---|---|---|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) | Frozen · native 0.1.54 · all attempts | 33 | 62.91 (40.33–76.26; n=33) | 825.88 (742.44–847.74; n=33) | 2.2 (0.6–15.3; n=33) | 18.75 (16.84–26.92; n=33) | 29.04 (28.13–29.21; n=33) |
| Gemma 4 (12B, Q4) | Frozen · native 0.1.54 · all attempts | 33 | 136.61 (109.89–159.16; n=33) | 1,437.51 (958.51–1,594.27; n=33) | 1.7 (0.5–16.2; n=33) | 16.33 (10.36–22.16; n=33) | 14.40 (13.49–21.08; n=33) |
Artifact reviews
Advisory means out of 10, with eligible artifact counts. Rubrics are separate and are not a combined model score. A dash means no eligible artifact was scored under that rubric. Missing artifacts have no score; excluded reviews do not enter these means. Judge: anthropic-cli/claude-sonnet-4-6. Review identity and limitations are recorded with the evaluation.
| Model | Period | HTML quality | No artifact | Excluded reviews |
|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) | Frozen · native 0.1.54 · all attempts | 8.39/10 (9 pieces) | 0 | 24 |
| Gemma 4 (12B, Q4) | Frozen · native 0.1.54 · all attempts | 6.97/10 (9 pieces) | 0 | 24 |
Tasks in this set: 11.
Office and knowledge work
| Model | Size | Tasks passed | Context | System memory |
|---|---|---|---|---|
| Gemma 4 (12B, Q4) (kv: f16) | 12B | 35/39 (90%) | 256K | 18.3 GB |
| Qwen 3.8 (27B, Q4) (kv: f16) | 27B | 33/39 (85%) | 139K | 22.2 GB |
Task performance
Task measurements across all attempted trials, including failures. Measurement methods and limits are stated in period labels and round notes. Values are per-trial medians (minimum–maximum; sample count), not admission probes. Wall time includes tools. System RSS and GPU memory are separate peak measurements; do not add their medians. Periods remain separate because hardware performance can change.
| Model | Period | Trials | Output (tok/s) | Prefill (tok/s) | Wall time (min) | Peak system RSS (GiB) | Peak GPU memory (GiB) |
|---|---|---|---|---|---|---|---|
| Gemma 4 (12B, Q4) | Frozen · native 0.1.54 · all attempts | 39 | 132.86 (112.51–153.97; n=39) | 1,319.42 (1,068.68–1,559.57; n=39) | 2.4 (0.5–31.4; n=39) | 18.26 (10.36–33.41; n=39) | 14.38 (13.57–14.61; n=39) |
| Qwen 3.8 (27B, Q4) | Frozen · native 0.1.54 · all attempts | 39 | 59.20 (41.81–67.67; n=39) | 809.47 (753.12–850.97; n=39) | 3.6 (0.8–42.0; n=39) | 22.20 (16.95–27.74; n=39) | 29.04 (28.22–29.30; n=39) |
Artifact reviews
Advisory means out of 10, with eligible artifact counts. Rubrics are separate and are not a combined model score. A dash means no eligible artifact was scored under that rubric. Missing artifacts have no score; excluded reviews do not enter these means. Judge: codex-inline/current-session. Review identity and limitations are recorded with the evaluation.
| Model | Period | craftbook-annotated-bibliography | craftbook-powerpoint-deck | craftbook-research-to-document | craftbook-spreadsheet-model | craftbook-week-plan | docblocks-theme-roundtrip | meeting-followup | wikipedia-research-brief | No artifact | Excluded reviews |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Gemma 4 (12B, Q4) | Frozen · native 0.1.54 · all attempts | 6.77/10 (3 pieces) | 7.70/10 (3 pieces) | 6.53/10 (3 pieces) | 3.50/10 (3 pieces) | 1.13/10 (3 pieces) | 4.10/10 (3 pieces) | 7.63/10 (3 pieces) | 2.97/10 (3 pieces) | 0 | 15 |
| Qwen 3.8 (27B, Q4) | Frozen · native 0.1.54 · all attempts | 6.70/10 (3 pieces) | 7.20/10 (3 pieces) | 6.67/10 (3 pieces) | 6.50/10 (3 pieces) | 1.70/10 (3 pieces) | 7.80/10 (3 pieces) | 7.77/10 (3 pieces) | 6.50/10 (3 pieces) | 0 | 15 |
Tasks in this set: 13.
Engineering work
| Model | Size | Tasks passed | Context | System memory |
|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) (kv: f16) | 27B | 19/30 (63%) | 139K | 22.8 GB |
| Gemma 4 (12B, Q4) (kv: f16) | 12B | 12/30 (40%) | 256K | 18.8 GB |
Task performance
Task measurements across all attempted trials, including failures. Measurement methods and limits are stated in period labels and round notes. Values are per-trial medians (minimum–maximum; sample count), not admission probes. Wall time includes tools. System RSS and GPU memory are separate peak measurements; do not add their medians. Periods remain separate because hardware performance can change.
| Model | Period | Trials | Output (tok/s) | Prefill (tok/s) | Wall time (min) | Peak system RSS (GiB) | Peak GPU memory (GiB) |
|---|---|---|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) | Frozen · native 0.1.54 · all attempts | 30 | 57.16 (44.17–69.52; n=30) | 800.20 (610.21–835.56; n=30) | 9.0 (1.1–55.8; n=30) | 22.82 (17.40–28.69; n=30) | 29.19 (28.22–30.07; n=29) |
| Gemma 4 (12B, Q4) | Frozen · native 0.1.54 · all attempts | 30 | 124.60 (113.31–149.03; n=30) | 1,041.47 (321.89–1,420.43; n=30) | 7.3 (0.8–61.6; n=30) | 18.76 (12.90–26.19; n=30) | 14.55 (13.57–14.86; n=30) |
Artifact reviews
Advisory means out of 10, with eligible artifact counts. Rubrics are separate and are not a combined model score. A dash means no eligible artifact was scored under that rubric. Missing artifacts have no score; excluded reviews do not enter these means. Judge: codex-inline/current-session. Review identity and limitations are recorded with the evaluation.
| Model | Period | craftbook-api-contract-review | craftbook-bug-fix-tdd | craftbook-code-review | craftbook-codemod-sweep | craftbook-deep-security-review | craftbook-refactor-module | No artifact | Excluded reviews |
|---|---|---|---|---|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) | Frozen · native 0.1.54 · all attempts | 7.17/10 (3 pieces) | 4.13/10 (3 pieces) | 7.17/10 (3 pieces) | 7.43/10 (3 pieces) | 7.70/10 (3 pieces) | 3.90/10 (3 pieces) | 0 | 12 |
| Gemma 4 (12B, Q4) | Frozen · native 0.1.54 · all attempts | 4.17/10 (3 pieces) | 4.30/10 (3 pieces) | 6.67/10 (3 pieces) | 5.47/10 (3 pieces) | 5.00/10 (3 pieces) | 3.37/10 (3 pieces) | 0 | 12 |
Tasks in this set: 10.
Complex workflows
| Model | Size | Tasks passed | Context | System memory |
|---|---|---|---|---|
| Gemma 4 (12B, Q4) (kv: f16) | 12B | 7/25 (28%) | 256K | 20.9 GB |
Task performance
Task measurements across all attempted trials, including failures. Measurement methods and limits are stated in period labels and round notes. Values are per-trial medians (minimum–maximum; sample count), not admission probes. Wall time includes tools. System RSS and GPU memory are separate peak measurements; do not add their medians. Periods remain separate because hardware performance can change.
| Model | Period | Trials | Output (tok/s) | Prefill (tok/s) | Wall time (min) | Peak system RSS (GiB) | Peak GPU memory (GiB) |
|---|---|---|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) | Frozen · native 0.1.54 · all attempts | 27 | 50.59 (44.40–70.60; n=27) | 762.01 (685.52–830.18; n=27) | 15.5 (1.0–54.4; n=27) | 26.42 (17.19–28.54; n=27) | 29.22 (28.69–29.52; n=27) |
| Gemma 4 (12B, Q4) | Frozen · native 0.1.54 · all attempts | 27 | 112.73 (98.44–155.81; n=27) | 1,031.92 (837.39–1,435.19; n=27) | 7.3 (0.8–30.1; n=27) | 20.86 (11.46–25.89; n=27) | 14.57 (14.03–15.09; n=27) |
Artifact reviews
Advisory means out of 10, with eligible artifact counts. Rubrics are separate and are not a combined model score. A dash means no eligible artifact was scored under that rubric. Missing artifacts have no score; excluded reviews do not enter these means. Judge: codex-inline/current-session. Review identity and limitations are recorded with the evaluation.
| Model | Period | craftbook-invoice-run | No artifact | Excluded reviews |
|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) | Frozen · native 0.1.54 · all attempts | 6.47/10 (3 pieces) | 0 | 24 |
| Gemma 4 (12B, Q4) | Frozen · native 0.1.54 · all attempts | 5.53/10 (3 pieces) | 0 | 24 |
Not published — some tasks could not be measured on this round: Qwen 3.8 (27B, Q4) (1 task(s) unmeasured)
Tasks in this set: 9.
October 28–30 R9700 comparison and review data
This note also covers 2026-10-30 · Windows · Ryzen 9 7950X3D · Radeon AI PRO R9700.
The Qwen 3.8 27B Q4 and Gemma 4 12B Q4 comparison completed all 258 trials across four task sets, with three repetitions per model and scenario. Qwen passed 104/125 eligible attempts; Gemma passed 81/127. Six raw exclusions remain recorded: three Windows sandbox failures and three model-authored gate configurations that the harness classified as infrastructure failures. Qwen's complex-work suite has no eligible custom-gate attempt, so that suite's headline is withheld while all its observations remain available.
Download the scorecard data and full run data with judging and performance. The companion includes per-trial fixed scores, both advisory judge ledgers, scenario and repetition results, latency, throughput, RAM/GPU measurements, postmortems and review audits. Known setup and evaluator problems remain visible beside the frozen results. Different review rubrics stay separate; misaligned A/B and incidental fixture-page opinions are excluded from the website's quality means and retained in the download.
A separate six-trial activation subset exposed the missing DocBlocks tools and produced two Qwen passes and no Gemma passes. It does not replace the original six setup failures. Binary repair messages were rejected before delivery; the postmortem distinguishes that framework limitation from initial schema and grounding errors. All trials and judging are complete.
Both models ran on Windows with a Radeon AI PRO R9700 and native 0.1.54. The task tables show engine-log measurements across all attempts; system RSS and GPU memory remain separate, and one GPU observation is missing. Download notes and validation explain all denominators, review selections, timing limitations and the frozen capability implementation's differences from its prose rubric.