Idol for mlx-serve (float16)
speridlabs/idol repackaged for mlx-serve on Apple
Silicon: one picture of an object becomes a 3D Gaussian splat, background removal included, no Python. POST /v1/3d/generations answers a GLB: a mesh of textured discs any viewer draws solid (Preview, Quick Look), or with
"geometry": "splat" the KHR_gaussian_splatting points. About 40 s for 50 steps and 262,144 Gaussians on an M-series
Ultra, 3.6 GB resident.
- The three transformers' Linear layers are float16; norms, embeddings and the RoPE phases are as released.
- Includes everything Idol reads, so nothing else downloads: BiRefNet (background removal, as released), the DINOv3
ViT-H+/16 encoder (float16; its licence is in
dinov3/LICENSE.md) and the FLUX.2 VAE encoder (vae/, from black-forest-labs/FLUX.2-klein-4B, Apache-2.0; the middle attention is 8-bit).
Licences
The Idol weights are Apache-2.0 (LICENSE, NOTICE). The bundled DINOv3 encoder is Meta's, under the DINOv3 License
(dinov3/LICENSE.md), which travels with any redistribution of it.
Use
mlx-serve --model-dir <folder holding this repo>, or pick Idol in the app's 3D pane.
About mlx-serve
A native Zig server for Apple Silicon. No Python, no cloud, no Electron. One binary: chat, images, video, music, speech and 3D on MLX, behind OpenAI- and Anthropic-compatible APIs, with a signed menu-bar app.
Converted by tests/convert_idol_weights.py in the mlx-serve repository.