Skip to content

ai-models.md

Latest commit

 

History

History
73 lines (51 loc) · 1.89 KB

ai-models.md

File metadata and controls

73 lines (51 loc) · 1.89 KB

AI Models

Included Models

All models are stored at /opt/volt/ai-models/ and are included in the ISO image.

Model File Size Format License
Qwen3 1.7B (Q8_0) ai-models/Qwen3-1.7B-Q8_0.gguf 1.7 GB GGUF Apache 2.0
Whisper Small ai-models/whisper/ggml-small.bin 465 MB GGML MIT
OuteTTS 0.2 (Q8_0) ai-models/outetts/OuteTTS-0.2-500M-Q8_0.gguf 512 MB GGUF Apache 2.0
WavTokenizer Large ai-models/outetts/WavTokenizer-Large-75-F16.gguf 124 MB GGUF MIT

Total AI model footprint: ~2.8 GB


Model Directory Structure

/opt/volt/ai-models/
├── Qwen3-1.7B-Q8_0.gguf
├── outetts/
│   ├── OuteTTS-0.2-500M-Q8_0.gguf
│   └── WavTokenizer-Large-75-F16.gguf
├── whisper/
│   └── ggml-small.bin
└── bin/
    └── whisper

Inference Runtimes

All models run via llama.cpp:

  • LLM: llama-server with Qwen3 GGUF — conversation, reasoning, code
  • STT: whisper.cpp with ggml-small — speech recognition
  • TTS: llama-tts with OuteTTS GGUF + WavTokenizer — text-to-speech

TTS Quality

OuteTTS 0.2 provides significantly better quality than the previous Piper TTS engine:

  • 82M parameter StyleTTS2 architecture
  • 24 kHz output sample rate
  • Near-human speech quality
  • WavTokenizer vocoder for audio decoding

Resource Reservations

Model VRM Quota DWP Workers Priority
Qwen3 (Orbit) 2048 MB 4 dedicated Critical
Whisper (STT) 512 MB Shared High
OuteTTS (TTS) 512 MB Shared High

Related


← Back: Documentation Index