The original model checkpoints are published by ACE-Step on Hugging Face. This repository does not change model architecture or training data; it repackages the released weights into the .bpk layout expected by Maolan's Rust/Burn runtime.

Included Models

The shared components are:

qwen3-encoder.bpk - Qwen3 embedding text encoder
acestep-dit.bpk - ACE-Step 1.5 turbo DiT
acestep-condition.bpk - condition stack and audio-code detokenizer
acestep-vae.bpk - Oobleck VAE decoder
silence_latent.bpk - silence timbre reference latent

The repository includes all turbo 5 Hz LM planner sizes supported by Maolan:

Planner	Weights	Config	Tokenizer
0.6B	`acestep-lm.bpk`	`lm_config.json`	`lm_tokenizer.json`
1.7B	`acestep-lm-1.7b.bpk`	`lm_config-1.7b.json`	`lm_tokenizer-1.7b.json`
4B	`acestep-lm-4b.bpk`	`lm_config-4b.json`	`lm_tokenizer-4b.json`

The 0.6B planner is the default. Larger planners require more memory and disk I/O, but may improve prompt planning.

Usage

Install or build Maolan Generate, then inspect the assets:

bash

maolan-generate \
  --inspect \
  --model acestep-turbo \
  --acestep-lm 0.6B \
  "dark rolling techno groove"

Generate audio:

bash

maolan-generate \
  --model acestep-turbo \
  --backend vulkan \
  --acestep-lm 1.7B \
  --bpm 128 \
  --key-scale "A minor" \
  --time-signature "4/4" \
  --length 10000 \
  --output loop.wav \
  "dark rolling techno groove"

To use a local checkout of this repository instead of downloading from the Hub:

bash

maolan-generate \
  --model acestep-turbo \
  --model-dir /path/to/ACE-Step-1.5-burn \
  --acestep-lm 4B \
  --output loop.wav \
  "cinematic synthwave instrumental"

Expected Runtime

maolan-generate --model acestep-turbo runs the turbo DiT schedule with 8 Euler steps. The generation pipeline is:

Embed the prompt and metadata with the Qwen3 text encoder.
Use the selected 5 Hz LM planner to generate FSQ audio semantic codes.
Use the turbo DiT to render 25 Hz audio latents from those codes.
Decode the latents to 48 kHz stereo audio with the Oobleck VAE.

Generation is currently instrumental. Lyrics or vocal conditioning are not implemented in Maolan's ACE-Step runtime.

File Layout

Maolan expects these files at the root of the model directory:

text

qwen3-encoder.bpk
qwen3_config.json
tokenizer.json
acestep-lm.bpk
lm_config.json
lm_tokenizer.json
acestep-lm-1.7b.bpk
lm_config-1.7b.json
lm_tokenizer-1.7b.json
acestep-lm-4b.bpk
lm_config-4b.json
lm_tokenizer-4b.json
acestep-dit.bpk
dit_config.json
acestep-condition.bpk
acestep-vae.bpk
vae_config.json
silence_latent.bpk

Conversion

These files were produced with the converter shipped in the Maolan source tree:

bash

cd generate
bin/convert_acestep.sh /path/to/out

The converter downloads the official ACE-Step 1.5 checkpoints and converts them to BurnPack. It handles the 4B planner as a sharded safetensors checkpoint.

Source code:

src/bin/acestep_convert.rs in Maolan Generate
bin/convert_acestep.sh in Maolan Generate

Source Checkpoints

The converted assets are derived from:

ACE-Step/Ace-Step1.5
ACE-Step/acestep-5Hz-lm-0.6B
ACE-Step/acestep-5Hz-lm-4B

The acestep-5Hz-lm-1.7B planner is distributed inside the ACE-Step/Ace-Step1.5 checkpoint tree.

License

The source ACE-Step releases are MIT licensed. This repository preserves that license for the converted model assets. Review the upstream ACE-Step model card and license before redistributing or using generated audio commercially.

Citation

If you use ACE-Step 1.5, cite the upstream project:

bibtex

@misc{gong2026acestep,
  title={ACE-Step 1.5: Pushing the Boundaries of Open-Source Music Generation},
  author={Junmin Gong and Yulin Song and Wenxiao Zhao and Sen Wang and Shengyuan Xu and Jing Guo},
  howpublished={\url{https://github.com/ace-step/ACE-Step-1.5}},
  year={2026},
  note={GitHub repository}
}

maolandaw/ACE-Step-1.5-burn

ACE-Step 1.5 BurnPack Assets for Maolan

Included Models

Usage

Expected Runtime

File Layout

Conversion

Source Checkpoints

License

Citation

maolandaw/
ACE-Step-1.5-burn