Benchmark page: Smarty DWQ 3-bit/4-bit vs served build on 500 GSM8K problems
Browse files- README.md +15 -0
- results.json +44 -0
README.md
CHANGED
|
@@ -22,6 +22,7 @@ A version counts as **better** only if it is faster or smaller *and* its answers
|
|
| 22 |
| Model | Mac | Base speed | Best version | Its speed | Its quality |
|
| 23 |
|---|---|---|---|---|---|
|
| 24 |
| Qwen3.5-4B (Smarty) | Apple M5 Pro (Mac mini) | 99.7 tok/s, quality 86.5% | base is still best | | |
|
|
|
|
| 25 |
| Qwen 3.8 27B | Apple M5 Pro (Mac mini) | 35.7 tok/s, quality 90.9% | base is still best | | |
|
| 26 |
| Qwen3.6-35B-A3B | Apple M3 Max, 36 GB (MacBook Pro) | 61.4 tok/s, quality – | base is still best | | |
|
| 27 |
| Qwen3.5-0.8B | Apple M5 Pro (Mac mini) | 120.0 tok/s, quality 7% | Qwen3.5-0.8B + MIT CSAIL's p50c weight edit (reproduced) | 120.0 tok/s | 33% |
|
|
@@ -50,6 +51,20 @@ Takeaway: the standard 4-bit build is still the best tradeoff. Speed tracks how
|
|
| 50 |
|
| 51 |
_Entered from the factory report as posted on the day; the full result files are on the Mac mini. Rows marked 'about' will be exact after the next run._
|
| 52 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
### Qwen 3.8 27B on Apple M5 Pro (Mac mini)
|
| 54 |
|
| 55 |
Measured 2026-11-02.
|
|
|
|
| 22 |
| Model | Mac | Base speed | Best version | Its speed | Its quality |
|
| 23 |
|---|---|---|---|---|---|
|
| 24 |
| Qwen3.5-4B (Smarty) | Apple M5 Pro (Mac mini) | 99.7 tok/s, quality 86.5% | base is still best | | |
|
| 25 |
+
| Qwen3.5-4B (Smarty), GSM8K math | Apple M5 Pro, 64 GB (Mac mini) | 99.7 tok/s, quality 89.0% | base is still best | | |
|
| 26 |
| Qwen 3.8 27B | Apple M5 Pro (Mac mini) | 35.7 tok/s, quality 90.9% | base is still best | | |
|
| 27 |
| Qwen3.6-35B-A3B | Apple M3 Max, 36 GB (MacBook Pro) | 61.4 tok/s, quality – | base is still best | | |
|
| 28 |
| Qwen3.5-0.8B | Apple M5 Pro (Mac mini) | 120.0 tok/s, quality 7% | Qwen3.5-0.8B + MIT CSAIL's p50c weight edit (reproduced) | 120.0 tok/s | 33% |
|
|
|
|
| 51 |
|
| 52 |
_Entered from the factory report as posted on the day; the full result files are on the Mac mini. Rows marked 'about' will be exact after the next run._
|
| 53 |
|
| 54 |
+
### Qwen3.5-4B (Smarty), GSM8K math on Apple M5 Pro, 64 GB (Mac mini)
|
| 55 |
+
|
| 56 |
+
Measured 2026-11-05.
|
| 57 |
+
|
| 58 |
+
| Version | Speed (tok/s) | vs base | Peak memory | Quality (GSM8K, 500 problems right) | Paired vs base | Verdict |
|
| 59 |
+
|---|---|---|---|---|---|---|
|
| 60 |
+
| **Standard 4-bit (what we serve)** | 99.7 | | 4.13 GB | 89.0% | | base |
|
| 61 |
+
| ↳ DWQ 3-bit (tuned to copy the full model) | 120.1 | +20% | 3.62 GB | 83.6% | +14 / -41 (p < 0.001) | worse: faster and smaller, but 2.5 to 8.3 points fewer right |
|
| 62 |
+
| ↳ DWQ 4-bit (tuned to copy the full model) | 100.4 | +1% | 4.13 GB | 88.4% | +20 / -23 (p = 0.76) | no gain: same speed, no real difference in right answers |
|
| 63 |
+
|
| 64 |
+
How it was measured: 500 GSM8K grade-school math problems (openai/gsm8k test, seed 0), the same for every version, served the way Smarty Pants serves Smarty: thinking off, greedy, up to 1,024 tokens. DWQ builds: mlx-lm DWQ, 1,024 samples, group size 64. A version is better only with more right answers at McNemar p < 0.05; faster counts only if the 95% interval of the accuracy change stays above -2 points. Script: scripts/model-factory/benchcheck.py.
|
| 65 |
+
|
| 66 |
+
Takeaway: neither DWQ build beats the one we serve. 3-bit DWQ is about 20% faster and 13% smaller but gets about 5 points fewer right; 4-bit DWQ is no different. The served build stays. We keep negative results on the page.
|
| 67 |
+
|
| 68 |
### Qwen 3.8 27B on Apple M5 Pro (Mac mini)
|
| 69 |
|
| 70 |
Measured 2026-11-02.
|
results.json
CHANGED
|
@@ -89,6 +89,50 @@
|
|
| 89 |
}
|
| 90 |
]
|
| 91 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 92 |
{
|
| 93 |
"id": "factory-qwen38-27b-apple-m5-pro-mac-mini",
|
| 94 |
"kind": "factory",
|
|
|
|
| 89 |
}
|
| 90 |
]
|
| 91 |
},
|
| 92 |
+
{
|
| 93 |
+
"id": "task-qwen35-4b-gsm8k-apple-m5-pro-mac-mini",
|
| 94 |
+
"kind": "task",
|
| 95 |
+
"order": 15,
|
| 96 |
+
"model": "Qwen3.5-4B (Smarty), GSM8K math",
|
| 97 |
+
"upstream": "Qwen/Qwen3.5-4B",
|
| 98 |
+
"mac": "Apple M5 Pro, 64 GB (Mac mini)",
|
| 99 |
+
"measured": "2026-11-05",
|
| 100 |
+
"quality_name": "Quality (GSM8K, 500 problems right)",
|
| 101 |
+
"test": "500 GSM8K grade-school math problems (openai/gsm8k test, seed 0), the same for every version, served the way Smarty Pants serves Smarty: thinking off, greedy, up to 1,024 tokens. DWQ builds: mlx-lm DWQ, 1,024 samples, group size 64. A version is better only with more right answers at McNemar p < 0.05; faster counts only if the 95% interval of the accuracy change stays above -2 points. Script: scripts/model-factory/benchcheck.py.",
|
| 102 |
+
"takeaway": "neither DWQ build beats the one we serve. 3-bit DWQ is about 20% faster and 13% smaller but gets about 5 points fewer right; 4-bit DWQ is no different. The served build stays. We keep negative results on the page.",
|
| 103 |
+
"rows": [
|
| 104 |
+
{
|
| 105 |
+
"id": "q4-served",
|
| 106 |
+
"label": "Standard 4-bit (what we serve)",
|
| 107 |
+
"role": "base",
|
| 108 |
+
"tps": 99.7,
|
| 109 |
+
"peak_gb": 4.13,
|
| 110 |
+
"quality": 0.89
|
| 111 |
+
},
|
| 112 |
+
{
|
| 113 |
+
"id": "dwq-3-g64",
|
| 114 |
+
"label": "DWQ 3-bit (tuned to copy the full model)",
|
| 115 |
+
"role": "version",
|
| 116 |
+
"tps": 120.1,
|
| 117 |
+
"peak_gb": 3.62,
|
| 118 |
+
"quality": 0.836,
|
| 119 |
+
"paired": "+14 / -41 (p < 0.001)",
|
| 120 |
+
"verdict": "worse",
|
| 121 |
+
"why": "faster and smaller, but 2.5 to 8.3 points fewer right"
|
| 122 |
+
},
|
| 123 |
+
{
|
| 124 |
+
"id": "dwq-4-g64",
|
| 125 |
+
"label": "DWQ 4-bit (tuned to copy the full model)",
|
| 126 |
+
"role": "version",
|
| 127 |
+
"tps": 100.4,
|
| 128 |
+
"peak_gb": 4.13,
|
| 129 |
+
"quality": 0.884,
|
| 130 |
+
"paired": "+20 / -23 (p = 0.76)",
|
| 131 |
+
"verdict": "no gain",
|
| 132 |
+
"why": "same speed, no real difference in right answers"
|
| 133 |
+
}
|
| 134 |
+
]
|
| 135 |
+
},
|
| 136 |
{
|
| 137 |
"id": "factory-qwen38-27b-apple-m5-pro-mac-mini",
|
| 138 |
"kind": "factory",
|