2026-10-30 · Windows · Ryzen 9 7950X3D · Radeon AI PRO R9700
A focused rerun on AMD Ryzen 9 7950X3D 16-Core Processor · AMD Radeon AI PRO R9700 · 64 GB: 2 models across 1 task set. 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 · focused subset · 3 trials per task · 2026-10-30 · gezel 22fec99111a5 · catalog 0.3.1
Separate six-trial DocBlocks activation condition, three repetitions per model. The supported binary expectedDeliverable contract exposed all 19 DocBlocks tools, changing the session/tool condition; it does not replace the 258-trial baseline. Qwen passed 2/3 and Gemma 0/3. Binary messageGezel repair was rejected before delivery, so these are not clean model-only failures or ignored nudges. Initial tool-schema and grounding errors are analyzed separately. Three Qwen reports were judged; three Gemma reports were absent. Original prompts, gates, generalist auto and native 0.1.54 were retained. See the separate condition postmortem in the R9700 download.
System memory counts this computer's own memory only. What the model keeps on the separate graphics card isn't included.
Office and knowledge work
| Model | Size | Tasks passed | Context | System memory |
|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) (kv: f16) | 27B | 2/3 (67%) | 139K | 22.1 GB |
| Gemma 4 (12B, Q4) (kv: f16) | 12B | 0/3 (0%) | 256K | 17.2 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) |
|---|---|---|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) | Separate activation condition · all attempts | 3 | 55.72 (54.59–56.53; n=3) | 780.47 (779.01–786.13; n=3) | 2.2 (1.9–7.0; n=3) | 22.13 (21.96–22.46; n=3) | 28.48 (28.48–28.49; n=3) |
| Gemma 4 (12B, Q4) | Separate activation condition · all attempts | 3 | 118.92 (117.91–124.89; n=3) | 1,199.86 (1,093.31–1,206.24; n=3) | 6.1 (5.9–6.8; n=3) | 17.23 (15.23–17.41; n=3) | 14.69 (14.69–14.70; n=3) |
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 | docblocks-theme-roundtrip | No artifact | Excluded reviews |
|---|---|---|---|---|
| Qwen 3.8 (27B, Q4) | Separate activation condition · all attempts | 5.43/10 (3 pieces) | 0 | 0 |
| Gemma 4 (12B, Q4) | Separate activation condition · all attempts | — | 3 | 0 |
Tasks in this set: 1.
October 28–30 R9700 comparison and review data
This note also covers 2026-10-28 · 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.