Head to head
d1-3B scores 78.4 and d1-omni-600M 65.2 on the System One Index. d1-3B is more accurate on 15 of 16 text datasets, d1-omni-600M on 1. Same cases, same request, same scorer.
| d1-3B | d1-omni-600M | |
|---|---|---|
| System One Index | 78.4 | 65.2 |
| English | 78.4 | 75.5 |
| Arabic | 78.3 | 53.4 |
| Vision Index | – | – |
| Calibration error | 2.4% | 9.8% |
| Median latency | – | – |
| Cost per 1,000 calls | – | – |
| d1-3B | d1-omni-600M | |
|---|---|---|
| AG News English | 88.5 | 87.5 |
| SST-2 English | 95.5 | 86.5 |
| SST-5 English | 56.0 | 47.0 |
| DAIR Emotion English | 63.5 | 85.0 |
| Banking77 English | 88.0 | 84.5 |
| BoolQ English | 85.0 | 74.0 |
| XNLI English English | 86.0 | 82.0 |
| Typed decisions English | 64.8 | 57.5 |
| XNLI French French | 78.7 | 65.3 |
| XNLI Arabic Arabic | 73.3 | 52.7 |
| MASSIVE scenario Arabic | 71.5 | 37.5 |
| MASSIVE intent Arabic | 83.5 | 52.0 |
| OSACT4 offensive Arabic | 84.5 | 77.5 |
| Tweet sentiment Arabic | 63.5 | 32.5 |
| MIRACL rerank Arabic | 82.5 | 55.0 |
| MIRACL pairwise Arabic | 89.0 | 66.5 |
Model pages: d1-3B · d1-omni-600M. Latency for API models is the median per call through OpenRouter, timed from one machine.