Releasing Depth-Anything-Edge: a 41 MB depth model that runs at 60 FPS on a Pi 5
A distilled monocular depth model with metric-stable outputs, a 0.32 AbsRel on KITTI, and a memory footprint that fits inside the Hailo-8L on a Pi HAT. Weights, recipes, and HIL traces, all open.
Depth-Anything-Edge is the second model we've shipped under our reference-distillation program, and the first one we'd recommend as a drop-in replacement for MiDaS-distilled-S in production. The headline: 41 MB on disk, 60 FPS on a Raspberry Pi 5 with a Hailo-8L HAT, 0.32 AbsRel on KITTI 2015 stereo eval. The full numbers — across all 17 chip platforms, in both single-stream and 4-stream configurations — are on the model page. This post is about what we changed to get there.
Depth-Anything-Edge is the second model we've shipped under our reference-distillation program, and the first one we'd recommend as a drop-in replacement for MiDaS-distilled-S in production. The headline: 41 MB on disk, 60 FPS on a Raspberry Pi 5 with a Hailo-8L HAT, 0.32 AbsRel on KITTI 2015 stereo eval. The full numbers — across all 17 chip platforms, in both single-stream and 4-stream configurations — are on the model page.
The interesting decisions were on the distillation side. We started from a frozen Depth-Anything-V2 ViT-B teacher and distilled into a hybrid stem + ViT-S/16 student. The thing that makes the student fit on edge silicon is not the parameter count — it's the rewriting of the attention path to fold the Q/K/V projections into a single per-head matmul, which lets the Hailo and CVflow compilers schedule it without ahead-of-time materialization.
We held two evaluations as gates. First, the AbsRel band on KITTI had to stay inside 0.34 — anything looser and we'd lose the metric-stable behavior practitioners rely on for robotics. Second, the latency at the target chip (Pi 5 + Hailo HAT) had to stay under 17 ms p50 — anything more and the model wouldn't fit inside the 60 FPS pipeline customers had built around the upstream model. We hit both, with margin.
Weights, the calibration set, the recipe, and the HIL traces are all open under Focsle-Research. There's a notebook on the GitHub repo that walks through the distillation step-by-step, and we'll be running an office-hours session next Thursday to take questions on it.