Skip to content

0.2.3

Choose a tag to compare

View all tags
@Saganaki22 Saganaki22 released this 05 Jul 06:37
v0.2.3
73a9c4c

Higgs Audio v3 Studio v0.2.3

Windows desktop maintenance release of Higgs Audio v3 Studio, a Rust/Tauri app for local Higgs Audio v3 TTS inference through a ported native C++/CUDA engine.

Downloads

Higgs Audio v3 Studio_0.2.3_x64-setup.exe

Windows NSIS installer. Use this for normal installation.

Higgs Audio v3 Studio_0.2.3_x64_en-US.msi

Windows MSI installer. Use this if you prefer MSI-based installation or deployment.

Higgs Audio v3 Studio_0.2.3_x64 Portable.zip/

Portable non-installer folder. Keep the .exe and resources/ folder together.

Requirements

  • Windows x64
  • NVIDIA GPU: 30xx, 40xx, or 50xx
  • CUDA 13 runtime/driver support

Recommended VRAM

Model Recommended VRAM
q4_k_m 8 GB
q5_k 9 GB
q6_k 10 GB
q8_0 12 GB
bf16 16 GB

What Is New In v0.2.3

  • Enabled F16 decode KV cache by default in the Higgs CUDA engine.
  • Reduced decode KV cache VRAM pressure compared with the previous F32 cache path.
  • Fixed the F16 KV cache crash caused by feeding F16 rows into ggml_set_rows.
  • Kept a safe F32 fallback path for diagnostics.
  • Disabled unsupported mixed K/V cache modes so they no longer hit CUDA flash-attention fatal errors.
  • Rebuilt the CUDA engine DLL for NVIDIA 30xx, 40xx, and 50xx targets.
  • Refreshed the bundled engine DLL in installer, MSI, and portable packages.
  • Updated Hugging Face manifest and SHA256SUMS.txt for the corrected engine DLL.
  • Kept the app version at v0.2.3.

What Is Included

  • Tauri desktop UI
  • Ported Higgs Audio v3 C++/CUDA engine DLL
  • Higgs config/tokenizer asset files
  • Model download UI
  • Whisper model selection/download UI
  • Voice clone workflow
  • Continue speech workflow
  • Multi-speaker workflow
  • Speaker Gallery
  • WAV/MP3 save support
  • Local API server
  • API streaming support
  • API examples for curl, Python, JavaScript, and PowerShell
  • API Command Centre log viewer
  • Bundled external-browser links for GitHub, releases, Hugging Face, and Whisper model selection

What Is Not Included

The large Higgs GGUF model weights are not bundled with the installer or portable build. Download them from the app or from the Hugging Face model repository.

Model files are hosted here:

https://huggingface.co/drbaph/Higgs-Audio-v3-Studio

Available Higgs Models

Model Description
q4_k_m Smaller model, recommended for 8 GB VRAM
q5_k Balanced K-quant model, recommended for 9 GB VRAM
q6_k Higher-quality K-quant model, recommended for 10 GB VRAM
q8_0 Recommended default, best quality/speed balance, recommended for 12 GB VRAM
bf16 Highest fidelity testing model, recommended for 16 GB VRAM

Installer Setup

  1. Download Higgs Audio v3 Studio_0.2.3_x64-setup.exe.
  2. Run the installer.
  3. Open Higgs Audio v3 Studio.
  4. Download/select the engine and model if needed.
  5. Load the engine, then load the model.
  6. Pick a workflow and generate audio.

Portable Setup

Download or copy the full Higgs Audio v3 Studio Portable/ folder.

Keep this structure intact:

Higgs Audio v3 Studio Portable/
├─ Higgs Audio v3 Studio.exe
└─ resources/
   ├─ engine/
   │  └─ audiocpp_engine.dll
   └─ higgs-assets/
      └─ higgs-audio-v3-tts-4b/
         ├─ config.json
         ├─ tokenizer.json
         ├─ tokenizer_config.json
         ├─ higgs_audio_v2_tokenizer_config.json
         └─ chat_template.jinja

Hugging Face Update Files

Upload these files to the Hugging Face model repository:

engines/audiocpp_engine.dll
manifest.json
checksums/SHA256SUMS.txt

Notes

  • The bundled engine DLL in this release is the corrected CUDA engine with F16 decode KV cache enabled by default.
  • The app downloads model weights separately into a user-writable model folder.
  • External links open in the system browser.
  • If the API Command Centre is popped out, it can remain visible while the main app is minimized to tray.
Assets 5
Loading