REAL LOCAL SUITES · CODING V2 + COMPREHENSIVE · SUITES NOT MIXED
LOCAL-FIRST · PUBLIC RESULTS
Model evidence, not model claims.
Standardized local runs with transparent hardware, parameters, raw answers, and reproducibility records.
Real results only
Scores below come from local captures on Apple M4 · 16 GB Unified Memory (macOS 27.0). Deleted models keep historical ranks with status Removed.
REAL LOCAL · CODING V2
Coding v2 leaderboard
temperature 0 · seed 42 · context 8192 · max output 1024 · thinking OFF · first-shot + one repair · hidden tests executed
Sort: Final ↓ · First ↓ · Gen tok/s ↓ (rank ties on Final)
| Rank | Model | First | Repair | Final | Gen tok/s | Prompt tok/s | Min Free | Swap | Status |
|---|---|---|---|---|---|---|---|---|---|
| 01 | official Qwen3.5-9B Q4_K_MQ4_K_M · llama.cpp 10450HB-20260912-C9B4A1 | 9 / 12 | 0/3 | 9 / 12 | 13.92 | 142.9 | 16% | 0.12 GB | Installed |
| 02 | Nemotron3 Nano 4B Q5_K_MQ5_K_M · llama.cpp 10450HB-20260912-N4B001 | 8 / 12 | 0/4 | 8 / 12 | 23.52 | 243.5 | 30% | 0 | Installed |
| 02 | gemma4-fable5Q5_K_M · ollamaHB-20260912-G4F8A2 | 8 / 12 | 0/4 | 8 / 12 | 10.81 | 104.6 | 8% | 2.63 GB | Installed |
| 02 | huihui Qwen3 abliterated 8BQ4_K_M · ollamaHB-20260912-HQ8B01 | 7 / 12 | 1/5 | 8 / 12 | 17.27 | 269.5 | 17% | 1.64 GB | Installed |
| 05 | Qwen3-4B-2507 HauhauCS Q6_KQ6_K · llama.cpp 10450HB-20260912-Q34B01 | 7 / 12 | 0/5 | 7 / 12 | 23.79 | 343.8 | 26% | 0 | Removed |
| 06 | Ministral-3-3B-Instruct-2512 Q5_K_MQ5_K_M · llama.cpp b10919HB-20260912-MN3B01 | 6 / 12 | 0/6 | 6 / 12 | 30.91 | 391.1 | 25% | 0 | Removed |
| 07 | HauhauCS Qwen3.5-4BQ4_K_M · ollamaHB-20260912-HQ4B01 | 4 / 12 | 1/8 | 5 / 12 | 26.86 | 371.3 | 24% | 0.10 GB | Installed |
| 08 | qwythos-9b-nsfwQ4_K_M · ollamaHB-20260912-QW9B01 | 4 / 12 | 0/8 | 4 / 12 | 16.78 | 239.7 | 17% | 1.79 GB | Installed |
| 08 | rocinante-nsfwQ4_K_M · ollamaHB-20260912-RO1201 | 3 / 12 | 1/9 | 4 / 12 | 13.21 | 208.9 | 9% | 1.08 GB | Installed |
REAL LOCAL · COMPREHENSIVE
Comprehensive benchmark
23-question fixed suite · temperature 0 · seed 42 · context 8192 · max output 128 · thinking OFF · rubric out of 80
Sort: Score ↓ · Gen tok/s ↓ (rank ties on Score)
Suite boundary
Coding v2 (12 tasks) and Comprehensive (80 points) use different rubrics. They are ranked separately and never mixed. Rank ties share the same number when the primary score (Final or Score) is equal.
| Rank | Model | Score | Max | Gen tok/s | Prompt tok/s | Min Free | Swap | Status |
|---|---|---|---|---|---|---|---|---|
| 01 | HauhauCS Qwen3.5-4BQ4_K_M · ollamaHB-20260911-HQ4BC80 | 49 | 80 | 26.09 | 240.8 | — | — | Installed |
| 02 | Qwen3.8-27B UD-IQ2_XXSUD-IQ2_XXS · llama.cpp 10450HB-20260911-Q3827B | 45.5 | 80 | 6.64 | 32.0 | 13% | 10.37 GB | Removed |
| 03 | qwythos-9b-nsfwQ4_K_M · ollamaHB-20260911-QW9B80 | 42.5 | 80 | 16.82 | 138.1 | 18% | 1.61 GB | Installed |
| 04 | gemma4-fable5Q5_K_M · ollamaHB-20260912-G4F80 | 41.5 | 80 | 6.87 | 75.4 | 8% | 2.27 GB | Installed |
| 05 | official Qwen3.5-9B Q4_K_MQ4_K_M · llama.cpp 10450HB-20260911-O9B80 | 41 | 80 | 9.28 | 73.3 | 18% | 3.37 GB | Installed |
| 06 | huihui Qwen3 abliterated 8BQ4_K_M · ollamaHB-20260912-HQ8B80 | 40 | 80 | 19.88 | 192.4 | 18% | 0.81 GB | Installed |
| 07 | Nemotron3 Nano 4B Q5_K_MQ5_K_M · llama.cpp 10450HB-20260912-N4B80 | 39.5 | 80 | 22.28 | 186.4 | 33% | 0 | Installed |
| 08 | Qwen3-4B-2507 HauhauCS Q6_KQ6_K · llama.cpp 10450HB-20260912-Q34B80 | 39 | 80 | 23.60 | 309.6 | 29% | 0 | Removed |
| 09 | Ministral-3-3B-Instruct-2512 Q5_K_MQ5_K_M · llama.cpp b10919HB-20260912-MN3B80 | 38.5 | 80 | 30.91 | 391.1 | 25% | 0 | Removed |
| 09 | HauhauCS Qwen3.5-9B Q6_KQ6_K · llama.cpp 10450HB-20260911-HQ9B80 | 38.5 | 80 | 7.89 | 71.2 | 11% | 3.63 GB | Removed |
| 11 | ultra-heretic 9B Q6_KQ6_K · llama.cpp 10450HB-20260911-UH9B80 | 35.5 | 80 | 7.86 | 72.4 | 13% | 4.06 GB | Removed |
| 12 | HauhauCS Gemma-4-E4B IQ4_XSIQ4_XS · llama.cpp 10450HB-20260912-G4E80 | 34.5 | 80 | 29.95 | 198.5 | 21% | 0.43 GB | Removed |
| 13 | rocinante-nsfwQ4_K_M · ollamaHB-20260912-RO1280 | 26.5 | 80 | 10.45 | 102.7 | 8% | 2.50 GB | Installed |
REAL PERFORMANCE · ELIGIBLE ONLY
Warm medium generation median
NO ELIGIBLE PERFORMANCE RUN
Canary, mock, partial, cleanup-failed, insufficient-sample, and invalid-timing records are excluded.
MACHINE
Apple M4 · 16 GB Unified Memory
| CPU | RAM | OS | GPU | Notes |
|---|---|---|---|---|
| Apple M4 | 16 GB Unified | macOS 27.0 | Apple M4 / MTL0 | Browser + IDE + terminal kept open during runs. |
REGISTRY SNAPSHOT