Instructions to use mochiexists528/ace-step-mlx-weights with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use mochiexists528/ace-step-mlx-weights with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir ace-step-mlx-weights mochiexists528/ace-step-mlx-weights
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
ACE-Step MLX Weights (fp16)
MLX-compatible safetensors of ACE-Step 1.5 for on-device music generation on Apple Silicon. Used by the mochi records iOS / macOS app.
Repository contents
text_encoder/ Qwen3-Embedding-0.6B text encoder (fp16 safetensors)
dit/ ACE-Step v1.5 turbo DiT + silence_latent (fp16 safetensors)
vae/ ACE-Step Oobleck VAE decoder (fp16 safetensors)
tokenizer/ Qwen3 tokenizer files (config + vocab)
Phase 6 LM-conditioning
Phase 6 is optional. Baseline generation only needs the text encoder, DiT,
VAE, tokenizer, and silence latent files above. To enable Phase 6, pass a
shared planner snapshot with --lm-weights.
The recommended planner repos are:
mochiexists528/ace-step-mlx-planner-1.7b-q4for lower storage and memory usemochiexists528/ace-step-mlx-planner-1.7bfor fp16 desktop/reference runs
The fp16 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 repository is a derivative work. It does not introduce new model architectures or weights — only a dtype repack of the upstream artifacts.
Upstream models
| Component | Upstream | License |
|---|---|---|
| DiT, VAE | ACE-Step/Ace-Step1.5 | MIT |
| Text encoder (Qwen3-Embedding-0.6B) | Qwen/Qwen3-Embedding-0.6B | Apache 2.0 |
| Tokenizer | Qwen3 tokenizer | Apache 2.0 |
Modifications from upstream
Apache 2.0 §4(b) requires us to state changes. The only modifications to the weights themselves are:
- dtype repack: upstream tensors that ship as
bfloat16are cast tofloat16for MLX safetensors compatibility. MLX did not load bfloat16 from safetensors at the time of conversion. No values are quantized, pruned, or otherwise altered beyond this dtype cast (precision delta < 1 ULP fp16 from the upstream bf16 round-trip). - File layout: tensors are repacked into the directory structure shown above. Tensor names and shapes are unchanged. Phase 6 planner weights are packaged in separate shared planner repos, not duplicated in this model repo.
No model architecture, training data, or learned parameters are modified.
NOTICE (Apache 2.0 §4(d))
Carrying forward the upstream Qwen3 NOTICE:
Qwen3
Copyright (c) Alibaba Cloud / Qwen team
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
The optional ACE-Step 5Hz planner LM is a fine-tune of Qwen3-1.7B and inherits the Qwen3 NOTICE obligation. It is distributed through the shared planner repos listed above, not through this model repo. See upstream Qwen3 model card for the canonical NOTICE text and verify the latest version before re-redistribution.
ACE-Step MIT license
Carrying forward the upstream ACE-Step MIT license:
MIT License — ACE-Step 1.5
Copyright (c) ACE-Step contributors
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
Usage
hf download mochiexists528/ace-step-mlx-weights --local-dir ./models
Then in the mochi records app, point the weight paths at the local checkout or let the in-app downloader pull from this repo on first launch.
Trademark notice
"ACE-Step" and "Qwen" are names of their respective upstream projects. This repository is not endorsed by, sponsored by, or affiliated with either project. Use of the names is solely for accurate attribution of the underlying weights.
Quantized