---
license: apache-2.0
tags:
  - music-generation
  - ace-step
  - gguf
  - dit
  - diffusion
  - task-arithmetic
  - model-merging
base_model:
  - ACE-Step/ACE-Step-v1-5-XL-DiT-Base
  - ACE-Step/ACE-Step-v1-5-XL-DiT-SFT
  - ACE-Step/ACE-Step-v1-5-XL-DiT-Turbo
pipeline_tag: text-to-audio
library_name: gguf
---

# ACE-Step 1.5 XL — Task Arithmetic Merged DiT Models (GGUF)

Quantized GGUF versions of **task-arithmetic merged** ACE-Step v1.5 XL DiT models, ready for use with C++ inference engines.

## What Are These?

These are **DiT (Diffusion Transformer)** models created by merging the official ACE-Step v1.5 XL checkpoints using [task arithmetic](https://arxiv.org/abs/2212.04089). Task arithmetic blends the learned "task vectors" of two fine-tuned models at a controlled interpolation ratio, producing models that inherit qualities from both parents.

Each model blends two of the three official XL checkpoints:

| Model | Parent A | Parent B | Ratio (λ) | Character |
|---|---|---|---|---|
| `merge-sft-turbo-xl-ta-0.3` | XL-SFT | XL-Turbo | 0.3 | Mostly SFT with a touch of Turbo speed |
| `merge-sft-turbo-xl-ta-0.7` | XL-SFT | XL-Turbo | 0.7 | Mostly Turbo with SFT musicality |
| `merge-base-turbo-xl-ta-0.5` | XL-Base | XL-Turbo | 0.5 | Equal blend of Base and Turbo |

> **λ = 0** means pure Parent A, **λ = 1** means pure Parent B.

### Why Merge?

- **SFT × Turbo** blends combine SFT's strong lyric adherence and musical structure with Turbo's faster convergence and energy
- **Base × Turbo** blends bring Base's raw generative range together with Turbo's efficiency
- Different ratios let you dial the trade-off to taste — lower λ for more structure, higher λ for more speed

## Architecture

All three models share the **XL** architecture:

| Parameter | Value |
|---|---|
| Architecture | `AceStepConditionGenerationModel` |
| Hidden size | 2560 |
| Intermediate size | 9728 |
| Attention heads | 32 (8 KV heads, GQA) |
| Layers | 32 (alternating sliding + full attention) |
| Encoder hidden size | 2048 |
| Head dim | 128 |
| Context length | 32768 |
| Parameters | ~4.7B |
| `is_turbo` | `false` (uses base-mode scheduling) |

## Available Quantizations

Each model is provided in 5 quantization levels:

| Quantization | Size | Notes |
|---|---|---|
| BF16 | 9,516 MB | Full precision, reference quality |
| Q8_0 | 5,060 MB | Near-lossless, recommended for quality |
| Q6_K | 3,909 MB | Excellent quality, good VRAM savings |
| Q5_K_M | 3,364 MB | Great balance of quality and size |
| Q4_K_M | 2,851 MB | Smallest, some quality trade-off |

### File Listing

```
acestep-v15-merge-base-turbo-xl-ta-0.5-BF16.gguf
acestep-v15-merge-base-turbo-xl-ta-0.5-Q4_K_M.gguf
acestep-v15-merge-base-turbo-xl-ta-0.5-Q5_K_M.gguf
acestep-v15-merge-base-turbo-xl-ta-0.5-Q6_K.gguf
acestep-v15-merge-base-turbo-xl-ta-0.5-Q8_0.gguf

acestep-v15-merge-sft-turbo-xl-ta-0.3-BF16.gguf
acestep-v15-merge-sft-turbo-xl-ta-0.3-Q4_K_M.gguf
acestep-v15-merge-sft-turbo-xl-ta-0.3-Q5_K_M.gguf
acestep-v15-merge-sft-turbo-xl-ta-0.3-Q6_K.gguf
acestep-v15-merge-sft-turbo-xl-ta-0.3-Q8_0.gguf

acestep-v15-merge-sft-turbo-xl-ta-0.7-BF16.gguf
acestep-v15-merge-sft-turbo-xl-ta-0.7-Q4_K_M.gguf
acestep-v15-merge-sft-turbo-xl-ta-0.7-Q5_K_M.gguf
acestep-v15-merge-sft-turbo-xl-ta-0.7-Q6_K.gguf
acestep-v15-merge-sft-turbo-xl-ta-0.7-Q8_0.gguf
```

## Compatibility

These GGUF files are **DiT-only** — they replace the `--dit` argument in the inference pipeline. You still need the standard LM, text encoder, and VAE models alongside them.

### [acestep.cpp](https://github.com/ServeurpersoCom/acestep.cpp)

Drop any of these GGUF files into the `models/` directory and pass via `--dit`:

```bash
./build/ace-server \
    --host 0.0.0.0 --port 8090 \
    --lm models/acestep-5Hz-lm-4B-Q8_0.gguf \
    --embedding models/Qwen3-Embedding-0.6B-Q8_0.gguf \
    --dit models/acestep-v15-merge-sft-turbo-xl-ta-0.3-Q8_0.gguf \
    --vae models/vae-BF16.gguf
```

### [HOT-Step-CPP](https://github.com/scragnog/HOT-Step-CPP)

Place the GGUF files in the engine's `models/` directory. The model will appear in the DiT model dropdown in the web UI automatically.

### Required Companion Models

These DiT GGUFs must be used alongside:

| Component | Model | Notes |
|---|---|---|
| LM | `acestep-5Hz-lm-4B-Q8_0.gguf` | Audio code language model |
| Text Encoder | `Qwen3-Embedding-0.6B-Q8_0.gguf` | Caption encoder |
| VAE | `vae-BF16.gguf` | Audio decoder (always BF16) |

Companion models are available from [Serveurperso/ACE-Step-1.5-GGUF](https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF).

## Recommended Settings

Since these are **non-turbo** (`is_turbo: false`) merged models, they use base-mode scheduling:

| Parameter | Recommended |
|---|---|
| Inference steps | 60–100 |
| CFG scale | 3.0–7.0 |
| Guidance mode | `apg` or `cfg` |
| Duration | 30–180s |

> **Note:** Higher step counts are needed compared to turbo models. These models trade speed for quality and creative range.

## How These Were Made

1. **Source checkpoints:** Official ACE-Step v1.5 XL safetensors (Base, SFT, Turbo)
2. **Merge method:** Task arithmetic — `θ_merged = θ_A + λ(θ_B − θ_A)`
3. **Conversion:** Safetensors → GGUF BF16 via `convert.py` from acestep.cpp
4. **Quantization:** BF16 → Q4_K_M / Q5_K_M / Q6_K / Q8_0 via the acestep.cpp `quantize` tool

## License

These models inherit the license from the upstream ACE-Step v1.5 checkpoints. See the [ACE-Step repository](https://github.com/ace-step/ACE-Step) for details.

## Credits

- **ACE-Step** — Original model architecture and training by the ACE-Step team
- **acestep.cpp** — C++ inference engine and GGUF tooling by [Serveurperso](https://github.com/ServeurpersoCom/acestep.cpp)
- **HOT-Step-CPP** — Full-stack music generation app by [scragnog](https://github.com/scragnog/HOT-Step-CPP)
- **Task arithmetic merges** — Produced by scragnog