At matched compute on an RTX 5070 Ti, DEIMv2 is the more accurate family: it averages 0.8 mAP points higher and wins 4 of 5 nearest-parameter pairs. DEIM answers on speed, running 10.4% faster on average in PyTorch. DEIMv2 also extends further down in size, fielding variants below 1M parameters.
DEIM fields 5 measured variants; DEIMv2 fields 8. Both run the same COCO val2017 protocol at 640 px on an RTX 5070 Ti, so the two ladders align by parameter count.
| Model | Family | Params (M) | mAP@50-95 | FPS |
|---|---|---|---|---|
| DEIMv2-Atto | DEIMv2 | 0.5 | 2746.0 | 40.4 |
| DEIMv2-Femto | DEIMv2 | 1.0 | 3452.0 | 45.3 |
| DEIMv2-Pico | DEIMv2 | 1.5 | 4227.0 | 40.5 |
| DEIMv2-N | DEIMv2 | 3.6 | 4671.0 | 35.1 |
| DEIM-N | DEIM | 3.8 | 4679.0 | 37.0 |
| DEIMv2-S | DEIMv2 | 9.8 | 5300.0 | 24.3 |
| DEIM-S | DEIM | 10.3 | 5210.0 | 33.4 |
| DEIMv2-M | DEIMv2 | 18.4 | 5600.0 | 22.1 |
| DEIM-M | DEIM | 19.6 | 5549.0 | 28.4 |
| DEIM-L | DEIM | 31.2 | 5783.0 | 18.6 |
| DEIMv2-L | DEIMv2 | 32.5 | 5856.0 | 19.7 |
| DEIMv2-X | DEIMv2 | 51.2 | 6134.0 | 17.4 |
| DEIM-X | DEIM | 62.6 | 5961.0 | 15.0 |
Accuracy at matched compute
mAP is shown in percent. Pairing by nearest parameter count, DEIMv2 wins 4 of 5 pairs and averages 0.8 mAP points higher. At the top end DEIMv2-X reaches 61.3 mAP against DEIM-X at 59.6. DEIM holds only the smallest matched pair.
Speed
DEIM is the faster family at matched compute: 10.4% faster on average in PyTorch FP32. The gap is widest in the small and mid classes, where DEIM-S runs 33.4 FPS against DEIMv2-S at 24.3.
- DEIM
- Apache-2.0 (permissive)
- DEIMv2
- Apache-2.0 (permissive)
- Commercial use
- Both families ship under Apache-2.0
Which family to pick
Pick DEIMv2 when accuracy leads or when you need the smallest models: it wins most matched pairs and fields variants down to 0.51M parameters. Pick DEIM when throughput is the constraint, since it runs faster at matched compute. Both ship under Apache-2.0. See the per-variant specs on the DEIM and DEIMv2 model pages.
Every number on this page comes from the verified dataset: same 500-image COCO val2017 slice, conf 0.001, IoU 0.6, max 300 detections, pycocotools mAP, identical protocol across all hardware and runtimes. The full protocol is on the methodology page. To rerun this comparison with your own filters, open compare. Accuracy is measured on LibreYOLO retrained checkpoints; other weight sources can yield different values.