Benchmarks / Which Small AI Model Should You Run on a Mac? We Ran 312 Tests / Jobs
rag · 40 items
Ask my documents
Answer questions about a folder of 22 household documents, with citations, and refuse when the answer isn't there.
Results
| Model | Accuracy | Pass | Easy | Medium | Hard | Median latency | Tok/s | Job time |
|---|---|---|---|---|---|---|---|---|
| MiMo-V2.6 Distill 9B | 90.8 | 90.0% | 100.0% | 92.3% | 80.0% | 4.11 s | 40.2 | 2m 48s |
| Granite 4.2 8B | 86.3 | 82.5% | 100.0% | 84.6% | 66.7% | 4.52 s | 39.3 | 3m 37s |
| Mellum2.1 12B-A2.5B | 83.8 | 80.0% | 66.7% | 92.3% | 80.0% | 2.53 s | 97.3 | 2m 03s |
| MiniCPM5 2B | 81.7 | 75.0% | 100.0% | 84.6% | 46.7% | 1.38 s | 103.6 | 56.1 s |
| Ornith-1.5 9B | 78.9 | 77.5% | 91.7% | 92.3% | 53.3% | 3.33 s | 47.5 | 2m 20s |
| Ling-3.0-tiny (7.9B-A1.3B) | 76.6 | 70.0% | 66.7% | 69.2% | 73.3% | 1.09 s | 111.0 | 45.1 s |
| Spark-X2.5 4B | 70.9 | 62.5% | 66.7% | 76.9% | 46.7% | 2.19 s | 68.3 | 1m 36s |
| K2-Horizon 0.9B | 41.6 | 27.5% | 50.0% | 15.4% | 20.0% | 0.64 s | 200.7 | 31.7 s |
How it is scored
Answerable: 0.7 x key-fact coverage + 0.3 x citation score (gold docs cited, -0.15 per extra doc cited); a refusal scores 0. Unanswerable: 1.0 for a refusal that asserts none of the trap facts, else 0. Pass = score >= 0.8.
Temperature 0 · max 800 output tokens.
System prompt every model got
You answer questions about the user's personal documents. Use only the document excerpts provided in the user's message, never outside knowledge or guesses. Keep the answer short (one to three sentences). Cite every document you used with its tag exactly as shown in the excerpts, for example [doc:lease]. If the excerpts do not contain the answer, reply with exactly this sentence and nothing else: I can't find that in your documents.
Every item
Final score per item (0-100). 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.