Head to head
d3 edge scores 65.9 and d3 lite 74.9 on the System One Index. d3 edge is more accurate on 0 of 16 text datasets, d3 lite on 15, and 1 are tied. Same cases, same request, same scorer.
| d3 edge | d3 lite | |
|---|---|---|
| System One Index | 65.9 | 74.9 |
| English | 69.1 | 74.3 |
| Arabic | 62.2 | 74.7 |
| Vision Index | – | – |
| Calibration error | 18.7% | 7.1% |
| Median latency | – | – |
| Cost per 1,000 calls | – | – |
| d3 edge | d3 lite | |
|---|---|---|
| AG News English | 80.0 | 85.5 |
| SST-2 English | 80.5 | 86.0 |
| SST-5 English | 39.5 | 50.5 |
| DAIR Emotion English | 57.0 | 57.0 |
| Banking77 English | 90.0 | 91.0 |
| BoolQ English | 79.5 | 81.0 |
| XNLI English English | 84.0 | 90.7 |
| Typed decisions English | 42.3 | 52.6 |
| XNLI French French | 66.0 | 81.3 |
| XNLI Arabic Arabic | 59.3 | 72.7 |
| MASSIVE scenario Arabic | 41.0 | 58.0 |
| MASSIVE intent Arabic | 55.0 | 77.5 |
| OSACT4 offensive Arabic | 72.5 | 80.5 |
| Tweet sentiment Arabic | 61.5 | 67.0 |
| MIRACL rerank Arabic | 70.0 | 82.5 |
| MIRACL pairwise Arabic | 76.0 | 84.5 |
Model pages: d3 edge · d3 lite. Latency for API models is the median per call through OpenRouter, timed from one machine.