mochiexists528/ace-step-mlx-weights-xl

ACE-Step MLX Weights XL (fp16)

MLX-compatible fp16 package for ACE-Step 1.5 XL-turbo. This repo follows the same layout as mochiexists528/ace-step-mlx-weights, with the XL-turbo DiT under dit/ and the shared text encoder, VAE, and tokenizer under their usual directories.

Contents

text_encoder/   Qwen3-Embedding-0.6B text encoder
dit/            ACE-Step v1.5 XL-turbo DiT + silence latent
vae/            ACE-Step Oobleck VAE decoder
tokenizer/      Qwen3 tokenizer files

Phase 6 LM-Conditioning

Phase 6 is optional. Baseline XL generation only needs the text encoder, DiT, VAE, tokenizer, and silence latent files above. To enable Phase 6 LM-conditioning, pass a shared planner snapshot with --lm-weights.

The recommended planner repos are:

• mochiexists528/ace-step-mlx-planner-1.7b-q4 for lower storage and memory use
• mochiexists528/ace-step-mlx-planner-1.7b for fp16 desktop/reference runs

The fp16 XL DiT carries the FSQ projection and AudioTokenDetokenizer tensors inline in dit/model.safetensors, so it does not need a separate detokenizer.safetensors conditioning sidecar.

Source Attribution

This is a derivative repack of upstream ACE-Step and Qwen artifacts for MLX. The XL-turbo DiT and VAE come from ACE-Step 1.5. The text encoder, tokenizer, and planner base family come from Qwen3. Planner weights are packaged in the shared Phase 6 planner repos listed above, not duplicated in this model repo. Tensor names and architecture are not changed; upstream bfloat16 tensors are stored as fp16 for MLX safetensors compatibility.