ACE-Step1.5#

  • Model Name: ACE-Step1.5

  • Model Family: ace_step_1_5

  • Abilities: [‘text2music’]

  • Multilingual: True

  • Engine: PyTorch

  • Python: 3.11 or 3.12

  • License: MIT

Specifications#

The built-in model downloads the complete ACE-Step 1.5 checkpoint bundle from Hugging Face or ModelScope. The bundle contains the default acestep-v15-turbo DiT, vae, Qwen3-Embedding-0.6B, and acestep-5Hz-lm-1.7B. Xinference uses the official ACE-Step 1.5 Python API in a per-model virtual environment.

This initial integration intentionally supports only those bundled DiT and LM checkpoints. Standalone ACE-Step DiT, LM, and VAE combinations are not selected through config_path or lm_model_path yet.

ACE-Step supports CUDA, ROCm, Apple Silicon, Intel XPU, and CPU. Accelerator availability and performance depend on the installed system PyTorch build.

Launch the default DiT-only configuration:

xinference launch --model-name ACE-Step1.5 --model-type audio --model-engine PyTorch

The default avoids loading the 5Hz language model. To enable LM planning, metadata completion, and audio-code reasoning, load the bundled 1.7B LM:

xinference launch --model-name ACE-Step1.5 --model-type audio \
  --model-engine PyTorch --lm_model_path acestep-5Hz-lm-1.7B

The LM uses its PyTorch backend by default. offload_to_cpu, offload_dit_to_cpu, quantization, and compile_model can be supplied as launch options for supported hardware. lm_backend accepts pt, vllm, or mlx; vllm requires CUDA for native execution, while mlx targets Apple Silicon. Unsupported hardware falls back according to the upstream ACE-Step runtime.

See ACE-Step1.5 speech usage for request examples and parameter limits.