Models

OSS distillation of π0. Validated on tabletop manipulation; HIL sim runs published alongside.

Edge-distilled OpenVLA. The smallest practical VLA you can put on an AGX Orin without compromising the policy.

Depth-Anything-Edge

Fo’c’sle

41%

Distilled depth model that holds 0.32 AbsRel on KITTI while running 60 FPS on a Pi 5 + Hailo HAT. The reference release.

Cosmos-distilled lane-change & merge policy. Targeted at Thor; AGX Orin runs in degraded mode.

Compact VLM that comfortably runs on a Jetson Orin Nano. The default 'on-device assistant for cameras' pick.

Production highway-perception detector — pedestrians, cars, motorcycles, trucks, lane delimiters. Same checkpoint we ship in Tier-1 cars.

SAM-3-Distilled

Fo’c’sle

28%

Distilled SAM-3 for VLA stacks that need on-device segmentation prompts. 4× lighter than the reference.

Real-time DETR optimized for transformer-friendly NPUs. Wins on transformer-class silicon, struggles on legacy NPUs.

Action-conditioned V-JEPA for world-model planning at the edge.

DINO-v3-Distilled

Fo’c’sle

21%

Distilled DINO-v3 backbone for downstream perception. The encoder under most of our reference models.

Apple's reference vision-language embedding for mobile. The encoder behind a lot of on-device retrieval.

Ninth-generation YOLO with edge-first quantization recipe. The default workhorse for 1080p multi-stream perception.

Smaller lane-change policy distilled from the Meridian Tier-1 stack. Validated for cut-in, merge, and yield in HIL.

Streaming-first embedded ASR. Lower latency than Whisper-Tiny at the cost of language coverage.

Open-vocabulary detector. Text-prompt the classes at runtime; no retraining.

Stereo-depth model paired with Meridian Highway-Perception v3. Optimized for 1216×352 KITTI-class input.

YOLO-backbone segment-anything variant. Where industrial inspection lives now.

Florence-2 distilled for edge inspection use cases — caption, detect, segment from a unified head.

Vendor-tuned YOLO27 for the Hailo Dataflow Compiler. Ships as native HEF.

Mobile-class RT-2 distillation under DeepMind's research license.

TensorRT-optimized YOLO27 with mixed-precision graph for Orin & Thor. Ships as TRT engine plan + ONNX fallback.

Real-time multi-person pose. Competes head-to-head with MoveNet on edge silicon.

Edge fork of Whisper-Tiny with NPU-friendly attention rewriting. INT4 variant runs on the Pi alone.

Layout-aware document parser. Pairs with PaddleOCR-v5 for end-to-end PDF understanding on-device.

ZoeDepth-Mobile

Fo’c’sle

7%

Metric depth in real units. Calibrated for mobile-tier NPUs.

Latest PaddleOCR with edge-tuned recognition head. 80+ languages, single binary.

Tiny variant tuned for sub-watt accelerators. Where the Coral TPU still earns its place.

Pure-tracker associator; pairs with any detector. The reference real-time MOT.

Larger MobileCLIP. Better recall, twice the compute.

Mobile-class panoptic segmenter. The reference for surgical, factory and medical pipelines.

Single-pose lightning model. 30 FPS on a Coral. The pose 'YOLO-Nano'.

Reference face embedding for access control and retail re-ID.

Hand-keypoint reference. Powers the AR pipelines you see on every demo phone.

DeepSORT successor with appearance embedding. Better cross-camera reacquisition than ByteTrack.

HF-Edge distillation. Good multilingual coverage on the higher-power tier.

Distilled MiDaS small. Maintained reference for cross-vendor depth comparisons.

Tiny pedestrian re-ID. The default for cross-camera tracking on dashcam-class silicon.

Reference EfficientDet variant pinned at INT8 for legacy NPUs. Maintained, not advanced.