# LocalAI Ships depth-anything.cpp: A Dependency-Free C++/ggml Port of ByteDance's Depth Anything 3

_Tool · published 2026-06-22_

The LocalAI team has released depth-anything.cpp, a from-scratch C++17/ggml port of ByteDance's Depth Anything 3 model that performs monocular metric depth estimation and camera pose inference with no Python, PyTorch, or CUDA dependencies at inference time. The self-contained GGUF package ships as a 99 MB quantized file (q4_k) and is verified bit-exact against the original PyTorch implementation at correlation 1.0.

On CPU, the port runs 1.20× faster than PyTorch at f32 and 1.31× faster at q8_0 quantization, while also loading approximately 6.7× faster and consuming roughly half the RAM. Beyond depth maps, the library outputs per-pixel confidence, camera extrinsics and intrinsics, an optional sky mask, and a back-projected 3D point cloud with export to glb, COLMAP, and PLY formats. The full Depth Anything 3 model family—from ViT-S to the nested Giant+Large metric model—is supported, as is Depth Anything V2, all driven by metadata baked into the GGUF file.

## Sources
- [github.com](https://github.com/mudler/depth-anything.cpp)
---
Canonical: https://genbuzz.news/posts/localai-ships-depth-anything-cpp-a-dependency-free-c-ggml-port-of-bytedance-s-depth-anything-3