mlx-community/ACE-Step-1.5-turbo-bf16

Text-to-Audio • 2B • Updated 1 day ago • 17

Note The 2 B flow DiT + condition encoder + FSQ tokenizer / detokenizer + silence latent (bf16, 4.5 GB). Required.

mlx-community/ACE-Step-1.5-planner-1.7B-bf16

Text-to-Audio • 2B • Updated 1 day ago • 20

Note The 5 Hz planner LM, Qwen3-1.7B fine-tune (bf16, 3.5 GB; the package quantises it 8-bit at load for its int8 tier). The default planner.

mlx-community/ACE-Step-1.5-planner-0.6B-bf16

Text-to-Audio • 0.7B • Updated 1 day ago • 22

Note The lite planner (Qwen3-0.6B fine-tune, 1.3 GB): the 24 GB-Mac tier; PER 0.179 vs the 1.7B's 0.108–0.197 across draws.

mlx-community/ACE-Step-1.5-text-encoder-bf16

Text-to-Audio • 0.6B • Updated 1 day ago • 23

Note Qwen3-Embedding-0.6B in mlx-lm layout (Apache-2.0, 1.1 GB): the caption encoder and lyric embedding table. Required.

mlx-community/ACE-Step-1.5-vae-fp32

Text-to-Audio • 0.2B • Updated 1 day ago • 19

Note The Oobleck VAE, encoder + decoder, weight norm folded (fp32, 0.6 GB). Required; kept fp32 on every tier.