Head to head
Clef scores 81.6 and d3 nano 78.0 on the System One Index. Clef is more accurate on 11 of 16 text datasets, d3 nano on 5. Same cases, same request, same scorer.
| Clef | d3 nano | |
|---|---|---|
| System One Index | 81.6 | 78.0 |
| English | 80.6 | 77.9 |
| Arabic | 83.0 | 76.9 |
| Vision Index | 78.1 | – |
| Calibration error | 2.2% | 7.0% |
| Median latency | 310 ms | – |
| Cost per 1,000 calls | $0.116 | – |
| Clef | d3 nano | |
|---|---|---|
| AG News English | 90.5 | 88.0 |
| SST-2 English | 97.0 | 89.5 |
| SST-5 English | 51.5 | 55.5 |
| DAIR Emotion English | 58.5 | 59.5 |
| Banking77 English | 95.5 | 93.0 |
| BoolQ English | 91.0 | 88.0 |
| XNLI English English | 88.7 | 87.3 |
| Typed decisions English | 72.2 | 62.6 |
| XNLI French French | 80.0 | 86.0 |
| XNLI Arabic Arabic | 78.0 | 82.0 |
| MASSIVE scenario Arabic | 88.0 | 61.5 |
| MASSIVE intent Arabic | 89.5 | 81.0 |
| OSACT4 offensive Arabic | 89.5 | 79.0 |
| Tweet sentiment Arabic | 58.0 | 71.0 |
| MIRACL rerank Arabic | 89.0 | 77.5 |
| MIRACL pairwise Arabic | 89.0 | 86.5 |
| Clef | d3 nano | |
|---|---|---|
| SST-2, as an image | 95.5 | – |
| Arabic sentiment, as an image | 35.0 | – |
| Oxford-IIIT Pet | 82.5 | – |
| VQAv2 yes/no | 87.5 | – |
| CountBench | 90.0 | – |
Model pages: Clef · d3 nano. Latency for API models is the median per call through OpenRouter, timed from one machine.