Head to head
d3 lite scores 74.9 and d3 nano 78.0 on the System One Index. d3 lite is more accurate on 3 of 16 text datasets, d3 nano on 13. Same cases, same request, same scorer.
| d3 lite | d3 nano | |
|---|---|---|
| System One Index | 74.9 | 78.0 |
| English | 74.3 | 77.9 |
| Arabic | 74.7 | 76.9 |
| Vision Index | – | – |
| Calibration error | 7.1% | 7.0% |
| Median latency | – | – |
| Cost per 1,000 calls | – | – |
| d3 lite | d3 nano | |
|---|---|---|
| AG News English | 85.5 | 88.0 |
| SST-2 English | 86.0 | 89.5 |
| SST-5 English | 50.5 | 55.5 |
| DAIR Emotion English | 57.0 | 59.5 |
| Banking77 English | 91.0 | 93.0 |
| BoolQ English | 81.0 | 88.0 |
| XNLI English English | 90.7 | 87.3 |
| Typed decisions English | 52.6 | 62.6 |
| XNLI French French | 81.3 | 86.0 |
| XNLI Arabic Arabic | 72.7 | 82.0 |
| MASSIVE scenario Arabic | 58.0 | 61.5 |
| MASSIVE intent Arabic | 77.5 | 81.0 |
| OSACT4 offensive Arabic | 80.5 | 79.0 |
| Tweet sentiment Arabic | 67.0 | 71.0 |
| MIRACL rerank Arabic | 82.5 | 77.5 |
| MIRACL pairwise Arabic | 84.5 | 86.5 |
Model pages: d3 lite · d3 nano. Latency for API models is the median per call through OpenRouter, timed from one machine.