Starstruk Logo MarkStarstruk

v0.1.6

snarktank/ralphv0.1.6Apr 3, 2026by ChuxiJ

AI Summary

Introduces the XL series, a larger 4B-parameter DiT decoder for higher audio quality that achieves the best scores across all 11 benchmark metrics. Includes new models (base, SFT, turbo) and extensive support for VRAM management, MLX on Apple Silicon, and training.

Key Highlights

  • Launch of XL series with 4B parameters and best benchmark scores
  • MLX Apple Silicon support for XL models
  • VRAM management and UI warnings for large model sizes
  • Training and LoRA support for XL models
  • Documentation available in 4 languages

New Features

  • Model download registry and VRAM management for XL
  • MLX Apple Silicon support for XL
  • UI VRAM warning for XL models
  • Training & LoRA support for XL
  • Documentation in 4 languages (en/zh/ja/ko)
  • Persist user preferences across sessions via localStorage

Full Release Notes

## 🎉 ACE-Step 1.5 XL (4B DiT) Release

We introduce the **XL series** — a larger 4B-parameter DiT decoder for higher audio quality. XL achieves **best scores across all 11 benchmark metrics**, surpassing all commercial and open-source models.

### New Models

| Model | Steps | CFG | HuggingFace |
|-------|:-----:|:---:|-------------|
| `acestep-v15-xl-base` | 50 | ✅ | [Link](https://huggingface.co/ACE-Step/acestep-v15-xl-base) |
| `acestep-v15-xl-sft` | 50 | ✅ | [Link](https://huggingface.co/ACE-Step/acestep-v15-xl-sft) |
| `acestep-v15-xl-turbo` | 8 | ❌ | [Link](https://huggingface.co/ACE-Step/acestep-v15-xl-turbo) |

- **Weights**: ~9GB (bf16) vs ~4.7GB for 2B
- **Min VRAM**: 12GB with offload, 20GB+ recommended
- **LM compatibility**: All LM models (0.6B/1.7B/4B) fully compatible

### Features
- Model download registry and VRAM management for XL (#989, #990)
- XL model code directories: xl_base, xl_sft, xl_turbo (#1001)
- MLX Apple Silicon support for XL (#1002)
- VRAM preflight check uses actual model size (#1003)
- UI VRAM warning for XL models (#1004)
- Documentation in 4 languages: en/zh/ja/ko (#1005)
- Training & LoRA support for XL (#1006)
- Deduplicated model config files (#1000)
- Default Gradio model changed to xl-turbo (#1009)
- Persist user preferences across sessions via localStorage (#1013)

### Bug Fixes
- Fix config_path lookup in VRAM guards and API /models endpoint (#1003)
- Fix MLX condition_embedder dimension mismatch for XL (#1002)
- Fix API auto-calculation for metadata params (duration/bpm) (#1023)
- Fix user preference restore regressions (#1025, #1026)
- Fix GenerationParams crash when None passed for numeric fields (#1027)
- Fix AAC/M4A source audio "invalid source audio" error in remix/extract (#1028)
- Fix empty values in restore_preferences causing Gradio crash on load (#1030)
- Fix cover/repaint using caller-supplied duration instead of src_audio duration (#1031, #1033)
- Fix cover/repaint/extract missing error when source audio not provided (#1035)
- Fix cover/repaint tasks with audio codes but no src_audio (#1036)

**Full Changelog**: https://github.com/ace-step/ACE-Step-1.5/compare/v0.1.5...v0.1.6