Bonframe LightVAE MLX BF16 decoders

BF16 MLX-layout decoder exports derived from lightx2v/LightVAE, upstream revision 4925ea978147b9de492d7a5fd4f9c6d1854ca1cc.

These are decoder-only weights for the Wan2.2 latent space. They require the corresponding Bonframe MLX LightVAE runtime; a standard Wan decoder cannot load these temporal low-rank factors directly. They are not a complete video generation or lip-sync model.

Directory	Weight bytes	Weight SHA-256
lite	304896412	5704b1bd835d143a9fc4b74dd322960bcdd61d349cc6ac440083fda0b9309598
pro	692371324	3bc7cf947363f18237c3fc2d9fe1a7083965c5bbe4e5dea779e1a10fb9c5a466

Temporal factors are folded on CPU in FP32 using the upstream implementation, then stored in BF16 with MLX convolution layouts. No additional integer quantization is applied. Config files describe the fixed architecture and source hashes. Reference NPZ files contain deterministic single- and multi-window Torch outputs for numerical validation.

Upstream source and licensing notices remain applicable. See UPSTREAM_NOTICE.txt and the upstream project for terms. This repository does not grant rights beyond those of the original assets.

Downloads last month: -; Downloads are not tracked for this model. How to track

MLX

Hardware compatibility

Quantized

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support