Seventeen configurations.
One 16GB GPU.

Quality, speed, and headroom are different choices.

These are workload-specific choices. The leader's gap to HauhauCS Q2_K_P is only 0.09 points, and the runner-up has a long-context failure. Fixed-seed repeats do not establish statistically significant ranking differences.

One ranking cannot describe every workload.

Bars use an absolute zero baseline. MTP entries are violet; the leading value is green. Timing is median request wall time with prompt-cache reuse, and output lengths vary.

Full results

Balanced index = 90% practical categories + 10% HumanEval+. Practical-only score renormalizes those categories to 100. Ties retain the same rank. Repository links resolve to the exact tested revision.

All 1,224 practical responses contained final text. Length finishes flag incomplete API completion, not necessarily a content failure. Six configurations had three length finishes each, for 18/1,224 responses. The revised rubric caps unfinished responses at 80.

MTP reduced measured response time in all four pairs.

These are base/MTP artifact comparisons with adaptive settings and differing output lengths, not identical-file on/off experiments. No HauhauCS custom FastMTP patch or sidecar was used.

Role leaders under the saved weighting

Every tied leader is listed. These are weighted practical/coding preferences, not separate evaluations. A role score does not override a failed long-context response.

Same local hardware. Versioned protocol.

One NVIDIA RTX 5070 Ti 16GB (16,303 MiB reported capacity), AMD Ryzen 5 5600G, llama.cpp 0.6.0 at commit d81235049384534c167caea52b85a694f6103d14. Each configuration ran three fixed and three adaptive dossier requests; 24 practical tasks three times; and one candidate on each of 164 HumanEval+ problems. Academic benchmark suites were not included.

Practical and HumanEval completion caps were 2,048 tokens, with a 512-token reasoning budget. The core dossier cap remained 3,072. Practical scoring revision 2.1 checks numeric equivalents, structured tool calls, schedule feasibility, negation, and selected robustness requirements. Legacy literal scores stay preserved as diagnostics and are not this leaderboard.

Context: the dossier exercised 57,135 prompt tokens at a 73,728-token allocation. Separate short-prompt allocation probes reached the per-model values below. The 14-marker dossier check is limited evidence; it is not validation of recall across 131K tokens.

Timing: three-request medians include cached prompt reuse. Request one processes the full dossier; requests two and three reuse roughly 57,131 prompt tokens. Faster wall time can reflect shorter output as well as faster generation. Empty final answers are excluded from successful-speed rankings.

Memory: headroom is the minimum sampled free GPU memory across adaptive repetitions; it is GPU-wide and includes runtime/KV allocations. File size is a separate metric.

Cost: $0 in cloud GPU rental for this local study. Electricity, hardware amortization, bandwidth and historic quant creation were not metered. This is not a claim of zero total project cost. The full study took 14h 20m wall time; the 14-entry queue took 11h 53m, including download, collection, scoring and the recovered interruption. Earlier audit/review work is excluded.

Keep the conclusions inside the evidence.

• Balanced index is 90% practical rubric scores and 10% HumanEval+. Long-context quality and speed are separate and do not enter that index.
• HauhauCS Q2_K_P returned empty final answers in all six long-context fixed/adaptive runs, despite API stop finishes. Its 6.90s adaptive median is excluded from successful-speed rankings; its saved 97.61 balanced score is preserved.
• Three fixed-seed repetitions test repeatability; they are not independent trials or confidence intervals. Small score gaps and ties at two-decimal precision are provisional.
• Timing medians include substantial prompt-cache reuse in repetitions 2 and 3. The first full-dossier request is reported separately. Different output lengths also affect wall time.
• Allocation probes use a short prompt. A successful 131072-token allocation is not a 131K recall test. The exercised dossier prompt was 57135 tokens at a 73728-token allocation.
• Retrieval hits use 14 literal markers within a synthetic dossier. They do not establish general long-context recall or overall intelligence.
• Practical scoring revision 2.1 uses deterministic content/structure checks. It has ceiling effects and incomplete semantic coverage; no general-purpose judge model was used.
• The isolated EvalPlus 0.3.1 scorer image was rebuilt after the original image disappeared. Six late configurations use the replacement; Python base and EvalPlus version match, but transitive dependencies were not fully locked. Per-model image IDs and recovery provenance are included.
• Practical profiles are frozen from the anchor audit or historical selections; the four new embedded-MTP models use ctx73728/ubatch256/MTP2. Core tuning remains adaptive. Settings and quant formats differ by configuration.
• HauhauCS uses standard embedded MTP under pinned llama.cpp; the custom FastMTP sidecar and patch were not tested. Uncensored labels are repository labels; this suite did not measure refusal or safety behavior.
• This study used one local RTX 5070 Ti 16GB. Historical report3 used RTX 5060 Ti and older protocol/runtime; do not combine scores or infer runtime-only causes from those differences.
• All model names and repository descriptions are retained as published identifiers. The report independently evaluates the downloaded pinned artifacts, not vendor speed claims.

All generation archives and evaluation manifests were checked before building this report. The original three cached GGUFs remain preserved; all fourteen temporary quants were removed after verified collection and scoring. The scorer-image reconstruction is recorded per affected configuration. No previous scores were overwritten.