EdgeFirst · Model Zoo / models · platforms · benchmarks · validation
The EdgeFirst Model Zoo
Multi-platform · multi-model · fully validatedThe EdgeFirst Model Zoo is a multi-platform, multi-model collection of production-ready vision models with full profiling and validation results for every published artifact. Each model is trained, exported, quantized, and compiled in EdgeFirst Studio, then validated on the actual deployment hardware — accuracy and timing are measured by the same pipeline that runs the deployed model, never a separate benchmark configuration.
On-target measurement is performed by the
EdgeFirst Profiler:
it runs full inference on the device through each runtime's native delegate, captures per-image
predictions as Parquet (inference results) and a Perfetto trace (.pftrace, detailed
per-stage timing), and ships both to EdgeFirst Studio for processing and publishing. Every
accuracy or timing number on this page links back to a public Studio validation session you can
inspect yourself. The Profiler is built on our open-source stack, including the
EdgeFirst HAL —
see github.com/EdgeFirstAI.
How the numbers are produced
.pftrace to Studio for scoring and publication.EdgeFirst Perception Index Report
Want the full story behind these numbers? The EdgeFirst Perception Index (EFPI) is a comprehensive technical report covering how we collected these benchmarks and validation results, how the models were optimized for each platform, and our insights and analysis of the results across the embedded hardware landscape.
Model Matrix
COCO val2017 · 640×640 · one repo per family × task| Repo | Family | Task | Sizes | Nano mAP50-95 (ONNX) | Nano mAP50-95 (INT8) | Last modified |
|---|
Hardware Targets
Platforms the zoo compiles models forNano Detection — mAP comparison
COCO val2017, ONNX FP32 vs TFLite INT8mAP@0.5-0.95 — ONNX FP32 vs TFLite INT8
mAP@0.5-0.95 vs Compute (Nano GFLOPs)
Nano Segmentation — Mask mAP
YOLOv8 / YOLO11 / YOLO26, COCO val2017Mask mAP@0.5-0.95 (ONNX vs INT8)
Detection mAP vs Mask mAP (ONNX FP32)
Accuracy vs On-target Throughput
COCO val2017 · pipelined FPS measured on each platform by the EdgeFirst ProfilerEach point is a public validation session: color = platform, shape = model family (● YOLOv5 · ■ YOLOv8 · ▲ YOLO11 · ◆ YOLO26). X axis is sustained pipelined throughput on a log scale — the full image load → decode → preprocess → inference → postprocess pipeline, not bare inference time. ⚠ flagged sessions are shown at their measured accuracy.
Accuracy vs On-target Latency
End-to-end per-image latency measured by the EdgeFirst Profiler on deployment hardwareOn-target latency results are being collected across the platform matrix. See the EdgeFirst Profiler documentation for how profiling and validation are performed.
Public Validation Sessions
every row links to its public session on EdgeFirst Studio⚠ Model accuracy on this platform is below our expectations — we are investigating the results to make improvements.
Detection — Platform Coverage
✓ = validated on-target with a public Studio sessionInstance Segmentation — Platform Coverage
✓ = validated on-target with a public Studio sessionModel Family Progress
Ultralytics YOLO families — published repos & on-target validation coverageAdditional model families and platforms are added to the zoo as their on-target validation completes.
Platform Throughput Leaderboard
Top 20 fastest validated model × platform results · full image load → decode → preprocess → inference → postprocess pipeline
| # | Platform | Fastest model | FPS (median) | E2E latency (ms) |
|---|
Throughput is the measured pipelined rate — stages overlap across frames, so it exceeds 1000 ÷ end-to-end latency. Measured by the EdgeFirst Profiler.
Per-platform Leaderboards
Ranked by COCO val2017 mAP@0.5-0.95 as measured on each platform by the EdgeFirst Profiler. Each entry links to its public validation session on EdgeFirst Studio.
⚠ Model accuracy on this platform is below our expectations — we are investigating the results to make improvements.