Unified MemoryBenchmarks / Which Small AI Model Should You Run on a Mac? We Ran 312 Tests / Jobs
chat · 42 items
Rewrite and proofread
Fix grammar and spelling without rewording, change a message's tone without losing facts, and shorten text to a word limit.
Results
| Model | Accuracy | Pass | Easy | Medium | Hard | Median latency | Tok/s | Job time |
|---|---|---|---|---|---|---|---|---|
| MiMo-V2.6 Distill 9B | 90.0 | 64.3% | 92.9% | 78.6% | 21.4% | 3.12 s | 36.2 | 2m 23s |
| Ornith-1.5 9B | 88.2 | 59.5% | 71.4% | 71.4% | 35.7% | 2.61 s | 45.4 | 1m 58s |
| Spark-X2.5 4B | 88.0 | 45.2% | 64.3% | 57.1% | 14.3% | 1.76 s | 63.0 | 1m 18s |
| Granite 4.2 8B | 82.9 | 47.6% | 71.4% | 50.0% | 21.4% | 2.74 s | 42.6 | 2m 26s |
| Ling-3.0-tiny (7.9B-A1.3B) | 77.7 | 26.2% | 21.4% | 35.7% | 21.4% | 0.84 s | 117.9 | 37.5 s |
| Mellum2.1 12B-A2.5B | 72.7 | 28.6% | 64.3% | 21.4% | 0.0% | 1.14 s | 96.2 | 51.4 s |
| MiniCPM5 2B | 67.6 | 21.4% | 28.6% | 35.7% | 0.0% | 1.23 s | 81.2 | 54.5 s |
| K2-Horizon 0.9B | 49.1 | 4.8% | 0.0% | 14.3% | 0.0% | 0.43 s | 218.4 | 19.6 s |
How it is scored
Deterministic checks, weighted, times a content gate. Proofread: errors fixed (wrong form gone and a corrected form present) 0.55, no rewording or new misspellings 0.15, names/numbers kept 0.1, correct-but-unusual phrasing kept 0.1, length 0.1. Tone: facts kept 0.45, banned strings/ALL-CAPS absent 0.2, no new numbers 0.15, length ratio 0.1, actually rewritten 0.1. Shorten: under the word limit 0.35, facts kept 0.45, no new numbers 0.2 (banned 0.1 where set). Tone and shorten also return needs_judge (rubric.md). Pass = every deterministic check passes (and, for judged items, judge mean >= 6).
Temperature 0 · max 1024 output tokens.
System prompt every model got
You are a careful writing assistant. When asked to proofread, rewrite, or shorten a text, reply with only the resulting text: no preamble, no explanation, no list of changes, no surrounding quotes, and no Markdown code fences.
Every item
Final score per item (0-100), which for this job blends the scripted checks with the judge panel's mark, as in the published results. Click any row to see the prompt, the expected answer and every model's reply. Columns sort.
✗ marks a fail (below the job's pass bar). A dash means no valid result.