YOLO26 Benchmark Results Note: This file is auto-generated by scripts/benchmark_yolo26_latency.py. To refresh these results, run: uv run python scripts/benchmark_yolo26_latency.py
Historical Note: This benchmark was originally created when comparing YOLO26 against RT-DETR. YOLO26 is now the sole detector in the system.
Date: 2026-01-26 06:30:35 GPU: NVIDIA RTX A5500 VRAM: 24576 MiB CUDA Version: 12.4 Driver Version: 550.54.14
Executive Summary This benchmark evaluates YOLO26 variants (nano, small, medium) for home security monitoring use cases.
Key Findings 640x640: Fastest model is yolo26n at 44.83ms (22.3 FPS)1280x1280: Fastest model is yolo26n at 129.33ms (7.7 FPS) Model Overview Model Parameters Size Architecture Task yolo26n 2.57M 5.3 MB YOLO26 Detection yolo26s 10.01M 19.5 MB YOLO26 Detection yolo26m 21.90M 42.2 MB YOLO26 Detection yolo26-v2-r101vd 76.56M 306.2 MB YOLO26 Detection
Latency Comparison (Single Image, Batch Size = 1) Lower is better. All values in milliseconds (ms).
640x640 Resolution Model Mean Std Min Max P50 P95 P99 yolo26n 44.83 1.38 41.17 47.71 44.94 46.77 47.57 yolo26s 73.74 2.01 68.82 78.59 74.33 75.62 77.56 yolo26m 138.84 1.68 134.87 144.32 138.95 141.46 143.50 yolo26-v2-r101vd 445.41 7.56 427.41 473.87 445.25 456.66 468.72
1280x1280 Resolution Model Mean Std Min Max P50 P95 P99 yolo26n 129.33 2.54 126.45 144.49 128.87 131.71 138.32 yolo26s 257.93 5.70 249.09 274.84 257.20 266.98 273.84 yolo26m 620.16 6.57 607.12 633.93 619.58 631.70 633.60 yolo26-v2-r101vd 429.76 8.13 414.90 444.05 428.76 443.18 443.95
Latency Breakdown (Batch Size = 1, 640x640) Shows preprocessing, inference, and postprocessing times.
Model Preprocess (ms) Inference (ms) Postprocess (ms) Total (ms) yolo26n 0.79 40.90 0.10 44.83 yolo26s 0.77 70.63 0.11 73.74 yolo26m 0.87 135.88 0.11 138.84 yolo26-v2-r101vd 3.85 441.14 0.42 445.41
Throughput Comparison (FPS) Higher is better. FPS = Frames Per Second.
640x640 Resolution Model Batch 1 Batch 4 Batch 8 Batch 16 yolo26n 22.3 32.7 38.3 32.1 yolo26s 13.6 17.6 15.8 13.5 yolo26m 7.2 6.5 5.6 5.3 yolo26-v2-r101vd 2.2 1.6 1.3 1.2
1280x1280 Resolution Model Batch 1 Batch 4 Batch 8 Batch 16 yolo26n 7.7 6.5 6.3 5.6 yolo26s 3.9 2.9 2.8 2.4 yolo26m 1.6 0.8 1.1 1.1 yolo26-v2-r101vd 2.3 1.4 1.2 1.1
VRAM Usage GPU memory requirements for deployment planning.
Model Resolution Batch Size VRAM Used (MB) VRAM Peak (MB) yolo26n 640x640 1 18377 18377 yolo26n 640x640 4 18376 18376 yolo26n 640x640 8 18376 18376 yolo26n 640x640 16 18376 18376 yolo26n 1280x1280 1 18376 18376 yolo26n 1280x1280 4 18376 18376 yolo26n 1280x1280 8 18376 18376 yolo26n 1280x1280 16 18431 18431 yolo26s 640x640 1 18427 18427 yolo26s 640x640 4 18427 18427 yolo26s 640x640 8 18426 18426 yolo26s 640x640 16 18430 18430 yolo26s 1280x1280 1 18439 18439 yolo26s 1280x1280 4 18485 18485 yolo26s 1280x1280 8 18339 18339 yolo26s 1280x1280 16 18347 18347 yolo26m 640x640 1 18339 18339 yolo26m 640x640 4 18343 18343 yolo26m 640x640 8 18343 18343 yolo26m 640x640 16 18364 18364 yolo26m 1280x1280 1 18586 18586 yolo26m 1280x1280 4 18934 18934 yolo26m 1280x1280 8 18463 18463 yolo26m 1280x1280 16 19530 19530 yolo26-v2-r101vd 640x640 1 19292 19292 yolo26-v2-r101vd 640x640 4 19643 19643 yolo26-v2-r101vd 640x640 8 18462 18462 yolo26-v2-r101vd 640x640 16 18336 18336 yolo26-v2-r101vd 1280x1280 1 18334 18334 yolo26-v2-r101vd 1280x1280 4 18336 18336 yolo26-v2-r101vd 1280x1280 8 18332 18332 yolo26-v2-r101vd 1280x1280 16 734 734
VRAM Requirements for Deployment Recommended VRAM allocation for different deployment scenarios:
Model Single Camera (640x640) Single Camera (1280x1280) Multi-Camera (4x 640x640) yolo26n 18377 MB 18376 MB 18376 MB yolo26s 18427 MB 18439 MB 18427 MB yolo26m 18339 MB 18586 MB 18343 MB yolo26-v2-r101vd 19292 MB 18334 MB 19643 MB
Recommendations For Real-Time Security Monitoring Use Case Recommended Model Rationale Low-power edge device yolo26n Lowest VRAM, fastest inference Standard security camera yolo26s Good balance of speed and accuracy High-resolution feeds yolo26m Better accuracy at higher resolutions Maximum accuracy yolo26-v2-r101vd Transformer architecture, best detection quality
Latency Targets for Security Applications Application Target Latency Recommended Configuration Real-time alert (<100ms) <50ms yolo26n/s at 640x640 Near real-time (<500ms) <200ms yolo26m at 640x640 or yolo26s at 1280x1280 Batch processing <1000ms Any model, optimize for throughput with larger batches
Benchmark Configuration Warmup iterations: 10Benchmark iterations: 100Batch sizes tested: 1, 4, 8, 16Resolutions tested: 640x640, 1280x1280 Notes All measurements taken on the same hardware under consistent conditions VRAM measurements include model weights and inference buffers Latency includes preprocessing, inference, and postprocessing YOLO26 uses end-to-end NMS-free inference which reduces postprocessing overhead YOLO26 uses SDPA attention when available for optimized inference