Configuration Parsing Warning:Config file tokenizer_config.json cannot be fetched (too big)

ACE-Step MLX Planner 1.7B Q4

MLX-compatible int4 package for the ACE-Step 5Hz planner LM used by optional Phase 6 LM-conditioning. This is the recommended planner package for device runs because it uses substantially less storage and memory than the fp16 planner.

The planner is shared by the base turbo and XL-turbo ACE-Step MLX model packages so it is not duplicated in every DiT repo.

Repository contents

config.json
manifest.json
model.safetensors
tokenizer.json
tokenizer_config.json
special_tokens_map.json
added_tokens.json
chat_template.jinja
merges.txt
vocab.json

Usage

Baseline generation does not need this repo. To enable Phase 6, pass the planner snapshot to the generator with --lm-weights.

For Q4 DiT variants, also pass the model repo's bundled detokenizer.safetensors as --lm-conditioning-checkpoint. For fp16 DiT variants, the FSQ projection and AudioTokenDetokenizer tensors are inline in dit/model.safetensors, so no detokenizer sidecar is needed.

hf download mochiexists528/ace-step-mlx-planner-1.7b-q4 --local-dir ./planner-q4

Source attribution

The planner LM is the ACE-Step 5Hz planner, a fine-tune in the Qwen3-1.7B family. Qwen3 is Apache 2.0 licensed. This package changes the on-disk layout and applies MLX affine int4 quantization to reduce storage and memory use.

NOTICE

This package carries the upstream Qwen3 NOTICE obligation. See the upstream Qwen3 model card for the canonical NOTICE text and verify the latest version before re-redistribution.

Downloads last month
33
Safetensors
Model size
0.3B params
Tensor type
BF16
·
U32
·
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support