- AnyDepth // 2601.02760
- MLX-OK
- WAVE 6
FRIGG
THE EFFICIENT
FRIGG achieves state-of-the-art monocular depth with only 14 million parameters — 85-89% fewer than standard DPT architectures. Built on DINOv3 with a Simple Depth Transformer, it uses quality-based data filtering to remove noisy training samples, achieving better accuracy with less data. Native Apple Silicon support via MLX makes it deployable on edge hardware without a GPU, running at under 100ms on M3 Max. This enables real-time depth perception on drones, mobile robots, and embedded systems — critical for edge robotics that cannot afford heavy GPU inference for basic depth perception.
MODULE STATUS: DEVELOPMENTParameters
14M
- DIVISION
- ANIMA
- WAVE
- W6
- DOMAIN
- DEPTH SYSTEMS
- WAVE 6 // ANIMA SUITE
- FOUNDATION — LIGHTWEIGHT DEPTH
HEAVY MODELS, LIGHT HARDWARE
Standard depth models use 100M+ params requiring heavy GPU. Edge devices cannot afford this.
The gap between model capability and hardware reality limits deployment.
WHAT FRIGG DELIVERS
FRIGG achieves SOTA monocular depth with only 14M parameters — 85-89% fewer than DPT. Built on DINOv3 with Simple Depth Transformer and native MLX support.
CAPABILITIES
- DINOv3 + Simple Depth Transformer — 14M params
- Quality-based data filtering
- Native Apple Silicon MLX — <100ms M3 Max
- Zero-shot camera transfer
WHY THIS IS HARD
Building FRIGG requires solving multiple coupled problems:
- 01Matching accuracy with 85-89% fewer parameters
- 02Quality curation at scale
- 03Cross-domain generalization with smaller model
- 04MLX optimization
FRIGG solves these through careful architecture design and rigorous validation.
PROOF, NOT PROMISES
Key metrics:
| METRIC | VALUE |
|---|---|
| Parameters | 14M |
| Reduction | 85-89% |
| Inference | <100ms |
| Backend | MLX Native |
WHAT'S BUILT TODAY
| COMPONENT | STATUS | NOTES |
|---|---|---|
| DINOv3 Backbone | COMPLETE | Pre-trained integrated |
| Depth Transformer | COMPLETE | SDT validated |
| Quality Filtering | COMPLETE | Auto noise removal |
| MLX Optimization | IN PROGRESS | Apple Silicon native |
| Core models | COMPLETE | Production-ready |
| API layer | IN PROGRESS | Edge REST API |
WHERE FRIGG DEPLOYS
- APP_01
DRONE PERCEPTION
Real-time depth on UAVs without GPU.
- APP_02
MOBILE ROBOTS
Lightweight depth for embedded platforms.
- APP_03
EDGE COMPUTING
Deploy on Apple Silicon for development.
UNDER THE HOOD
FOUNDATION: ANYDEPTH
- DINOv3 + Simple Depth Transformer — 14M params
- Quality-based data filtering
- Native Apple Silicon MLX — <100ms M3 Max
KEY INNOVATION
FRIGG achieves SOTA monocular depth with only 14M parameters — 85-89% fewer than DPT. Built on DINOv3 with Simple Depth Transformer and native MLX support.
DEPLOYMENT
- REST API
- Docker containerized
- Prometheus metrics
- Configurable backends
COMPUTE
- PRIMARY
- MLX-OK
- EDGE
- Optimized inference
- API
- REST + streaming
PAPERS
- [01]AnyDepth (2601.02760)