serveurperso/ace-step-1.5-gguf overview

{
  "metadata": {},
  "card_data": {
    "language": [
      "en",
      "fr",
      "zh",
      "ja",
      "ko",
      "de",
      "es",
      "it",
      "pt",
      "ru"
    ],
    "tags": [
      "music-generation",
      "text-to-music",
      "ggml",
      "gguf",
      "cpp",
      "ace-step",
      "diffusion",
      "flow-matching"
    ],
    "license": "mit",
    "base_model": [
      "ACE-Step/Ace-Step1.5"
    ],
    "frontmatter": {
      "language": [
        "en",
        "fr",
        "zh",
        "ja",
        "ko",
        "de",
        "es",
        "it",
        "pt",
        "ru"
      ],
      "tags": [
        "music-generation",
        "text-to-music",
        "ggml",
        "gguf",
        "cpp",
        "ace-step",
        "diffusion",
        "flow-matching"
      ],
      "license": "mit",
      "base_model": [
        "ACE-Step/Ace-Step1.5"
      ]
    },
    "hero_image_url": "",
    "summary": "Pre-quantized GGUF models for acestep.cpp, a portable C++17 implementation of the ACE-Step 1.5 AI Music Generation using GGML. Text + lyrics in, stereo 48kHz audio out. Runs on CPU, CUDA, Metal, Vulkan.",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlanguage:\n- en\n- fr\n- zh\n- ja\n- ko\n- de\n- es\n- it\n- pt\n- ru\ntags:\n- music-generation\n- text-to-music\n- ggml\n- gguf\n- cpp\n- ace-step\n- diffusion\n- flow-matching\nlicense: mit\nbase_model:\n- ACE-Step/Ace-Step1.5\n---\n\n# ACE-Step 1.5 GGUF\n\nPre-quantized GGUF models for [acestep.cpp](https://github.com/ServeurpersoCom/acestep.cpp), a portable C++17 implementation of the ACE-Step 1.5 AI Music Generation using GGML.\n\nText + lyrics in, stereo 48kHz audio out. Runs on CPU, CUDA, Metal, Vulkan.\n\n## Quick start\n\n```bash\ngit clone --recurse-submodules https://github.com/ServeurpersoCom/acestep.cpp\ncd acestep.cpp\n\npip install huggingface_hub\n./models.sh           # downloads Q8_0 turbo essentials (~7.7 GB)\n\nmkdir build && cd build\ncmake .. -DGGML_CUDA=ON\ncmake --build . --config Release -j$(nproc)\ncd ..\n\n./build/ace-server --models ./models --host 0.0.0.0 --port 8085\n```\n\nOpen http://localhost:8085 in your browser. The embedded WebUI handles everything:\nwrite a caption, set lyrics, generate, play, and download tracks.\n\nModels are loaded on demand and swapped automatically from the UI.\n\n<details>\n<summary>CLI tools (without the server)</summary>\n\n```bash\n# LM: generate lyrics + audio codes\n./build/ace-lm \\\n    --request /tmp/request.json \\\n    --lm models/acestep-5Hz-lm-4B-Q8_0.gguf\n\n# DiT + VAE: synthesize audio\n./build/ace-synth \\\n    --request /tmp/request0.json \\\n    --embedding models/Qwen3-Embedding-0.6B-Q8_0.gguf \\\n    --dit models/acestep-v15-turbo-Q8_0.gguf \\\n    --vae models/vae-BF16.gguf\n```\n\n</details>\n\n## Available models\n\n### Text encoder\n\n| File | Quant | Size |\n|------|-------|------|\n| Qwen3-Embedding-0.6B-BF16.gguf | BF16 | 1.2 GB |\n| Qwen3-Embedding-0.6B-Q8_0.gguf | Q8_0 | 748 MB |\n\nFrozen Qwen3 encoder (28 layers, 1024-dim). The DiT was trained end-to-end with this exact model. Its CondEncoder projection weights (1024 to 2048) are baked into every DiT checkpoint, so the Text-Enc is architecturally locked to 0.6B.\n\n### LM (Qwen3 causal, audio code generation)\n\n| File | Params | Quant | Size |\n|------|--------|-------|------|\n| acestep-5Hz-lm-4B-BF16.gguf | 4B | BF16 | 7.9 GB |\n| acestep-5Hz-lm-4B-Q8_0.gguf | 4B | Q8_0 | 4.2 GB |\n| acestep-5Hz-lm-4B-Q6_K.gguf | 4B | Q6_K | 3.3 GB |\n| acestep-5Hz-lm-4B-Q5_K_M.gguf | 4B | Q5_K_M | 2.9 GB |\n| acestep-5Hz-lm-1.7B-BF16.gguf | 1.7B | BF16 | 3.5 GB |\n| acestep-5Hz-lm-1.7B-Q8_0.gguf | 1.7B | Q8_0 | 1.9 GB |\n| acestep-5Hz-lm-0.6B-BF16.gguf | 0.6B | BF16 | 1.3 GB |\n| acestep-5Hz-lm-0.6B-Q8_0.gguf | 0.6B | Q8_0 | 677 MB |\n\nSmall LMs (0.6B/1.7B) only have BF16 + Q8_0 (too small for aggressive quantization).\nThe 4B LM does not have Q4_K_M (breaks audio code generation).\n\n### DiT (flow matching diffusion transformer)\n\n#### Standard (2B)\n\nAvailable for 7 variants: turbo, sft, sftturbo50, base, turbo-shift1, turbo-shift3, turbo-continuous.\n\n| Quant | Size per variant |\n|-------|-----------------|\n| BF16 | 4.5 GB |\n| Q8_0 | 2.4 GB |\n| Q6_K | 1.9 GB |\n| Q5_K_M | 1.6 GB |\n| Q4_K_M | 1.4 GB |\n\n#### XL (4B)\n\nAvailable for 4 variants: xl-turbo, xl-sft, xl-sftturbo50, xl-base.\n\n| Quant | Size per variant |\n|-------|-----------------|\n| BF16 | 9.3 GB |\n| Q8_0 | 5.0 GB |\n| Q6_K | 3.9 GB |\n| Q5_K_M | 3.3 GB |\n| Q4_K_M | 2.8 GB |\n\nTurbo: 8 steps. SFT/Base: 32-50 steps. SftTurbo50: weight merge provides a little more richness from the SFT while maintaining the low number of Turbo steps.\n\n### VAE\n\n| File | Size |\n|------|------|\n| vae-BF16.gguf | 322 MB |\n\nAlways BF16 (small, bandwidth-bound, quality-critical).\n\n## Pipeline\n\n```\nCompose (ace-lm):\n\n  Caption -> Qwen3 LM (0.6B/1.7B/4B) -> metadata + lyrics + audio codes (5Hz)\n\nSynthesize (ace-synth / ace-server):\n\n  Caption + Lyrics -> Text-Enc (Qwen3-Embedding-0.6B) -> CondEncoder\n  Audio codes 5Hz -> FSQ detokenizer (neural net in DiT GGUF) -> latents 25Hz\n  LoRA (optional) -> DiT (flow matching, Euler steps) -> latents 25Hz -> VAE decode -> WAV 48kHz\n\nCover modes (ace-server):\n\n  Source audio -> VAE encode -> latents 25Hz -> DiT context\n  Reference audio -> timbre conditioning\n```\n\nThe LM and DiT were co-trained on the same music data. The LM operates at 5Hz (each token = 200ms of music, vocabulary of 64000 learned codes) and builds the global musical structure autoregressively with creative sampling. The DiT takes over at 25Hz (one frame every 40ms) and uses flow matching to render the high-frequency details: timbre, transients, vocal articulation, stereo imaging.\n\nBoth stages support batching for parallel generation.\n\n## Music guide\n\nFor tips on writing effective prompts, understanding inference parameters, and getting the best results:\n\n- [A Musician's Guide](https://github.com/ace-step/ACE-Step-1.5/discussions/235) (non-technical, for music makers)\n- [Tutorial](https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/Tutorial.md) (design philosophy, architecture, hyperparameters)\n\n## Acknowledgements\n\nIndependent C++/GGML implementation based on [ACE-Step 1.5](https://github.com/ace-step/ACE-Step-1.5) by ACE Studio and StepFun. All original model weights are theirs.\n\nSFT/Turbo merge weights (standard 2B DiT) by [Aryanne](https://huggingface.co/Aryanne).\nSFT/Turbo merge weights (XL 4B DiT) by [jeankassio](https://huggingface.co/jeankassio).\n\n## Links\n\n- [acestep.cpp](https://github.com/ServeurpersoCom/acestep.cpp) - source code\n- [ACE-Step 1.5](https://github.com/ace-step/ACE-Step-1.5) - original Python implementation\n- [ACE-Step model hub](https://huggingface.co/ACE-Step) - original weights\n",
    "related_quantizations": []
  },
  "tags": [
    "gguf",
    "music-generation",
    "text-to-music",
    "ggml",
    "cpp",
    "ace-step",
    "diffusion",
    "flow-matching",
    "en",
    "fr",
    "zh",
    "ja",
    "ko",
    "de",
    "es",
    "it",
    "pt",
    "ru",
    "base_model:ACE-Step/Ace-Step1.5",
    "base_model:quantized:ACE-Step/Ace-Step1.5",
    "license:mit",
    "endpoints_compatible",
    "region:us"
  ],
  "likes": 43,
  "downloads": 60893,
  "gated": false,
  "private": false,
  "last_modified": "2026-04-14T17:47:28.000Z",
  "created_at": "2026-02-20T18:47:13.000Z",
  "pipeline_tag": "",
  "library_name": ""
}

{
  "_id": "6998ac31c055587af21a8186",
  "id": "Serveurperso/ACE-Step-1.5-GGUF",
  "modelId": "Serveurperso/ACE-Step-1.5-GGUF",
  "sha": "9b3707625776cc4cf775e9b12ab82f9fe48335ff",
  "createdAt": "2026-02-20T18:47:13.000Z",
  "lastModified": "2026-04-14T17:47:28.000Z",
  "author": "Serveurperso",
  "downloads": 60893,
  "likes": 43,
  "gated": false,
  "private": false,
  "pipeline_tag": "",
  "library_name": "",
  "siblings_count": 65
}