Clark Air · Sana 1.6B · MLX 2-bit

The Clark Air Sana 1.6B transformer quantized to ternary, stored in MLX 2-bit kernels for Apple Silicon — the first MLX port of Sana. 543 MB vs 3.21 GB FP16.

Generated end-to-end on an M5 Max: MLX 2-bit transformer driving the denoising loop, VAE decode.

Contents

• clark_air_sana_mlx_2bit.safetensors — MLX-quantized transformer (trunk ternary in MLX 2-bit slots at per-layer power-of-two groups; conditioning/projection layers kept high-precision).
• mlx_sana.py — the MLX runtime / loader.

Status

The transformer (the compressed core), verified to reproduce the FP16 model across a full 20-step generation (3.4% per-step). A full standalone MLX pipeline (scheduler + text encoder + VAE) is in progress; until then run the transformer with the rest of the Sana pipeline bridged from diffusers.

Usage

import mlx.core as mx
from mlx_sana import SanaMLX
sd = mx.load("clark_air_sana_mlx_2bit.safetensors")
model = SanaMLX(sd)  # auto-detects the packed 2-bit trunk
# noise = model(latent, encoder_hidden_states, timestep, enc_mask=mask)

License

Apache-2.0 © Clark Labs, Inc.