IzwiIzwi

Models

Izwi uses AI models for text-to-speech, speech recognition, and chat. This guide explains how to find, download, and manage models.

Available Models

Izwi supports several model families optimized for different tasks:

Text-to-Speech (TTS)

ModelSizeDescription
Qwen3-TTS-12Hz-0.6B-Base~2.3 GBFast, general-purpose TTS
Qwen3-TTS-12Hz-0.6B-CustomVoice~2.3 GBTTS with voice cloning support
Qwen3-TTS-12Hz-1.7B-Base~4.2 GBHigher quality TTS
Qwen3-TTS-12Hz-1.7B-CustomVoice~4.2 GBHigher quality with voice cloning
Qwen3-TTS-12Hz-1.7B-VoiceDesign~4.2 GBVoice design from descriptions
LFM2-Audio-1.5B~2.8 GBLiquid AI audio model
LFM2.5-Audio-1.5B~3.0 GBLiquid AI audio model (v2.5)

Speech Recognition (ASR)

ModelSizeDescription
Qwen3-ASR-0.6B~1.8 GBFast speech-to-text
Qwen3-ASR-1.7B~4.4 GBHigher accuracy transcription
Parakeet-TDT-0.6B-v2~4.6 GBNVIDIA Parakeet ASR
Parakeet-TDT-0.6B-v3~9.4 GBNVIDIA Parakeet ASR (latest)

Speaker Diarization

ModelSizeDescription
diar_streaming_sortformer_4spk-v2.1~0.5 GBNVIDIA Sortformer, up to 4 speakers

Chat

ModelSizeDescription
Qwen3-0.6B-4bit~0.8 GBCompact Qwen3 chat model
Qwen3-1.7B~3.8 GBLarger Qwen3 chat model
Gemma-3-1b-it~2.0 GBGoogle Gemma 3 1B Instruct
Gemma-3-4b-it~8.0 GBGoogle Gemma 3 4B Instruct

Specialized

ModelSizeDescription
Qwen3-ForcedAligner-0.6B~1.7 GBWord-level audio alignment
Voxtral-Mini-4B-Realtime-2602~7.5 GBMistral realtime audio (coming soon)

Downloading Models

Via CLI

# List all available models izwi list # Download a model izwi pull qwen3-tts-0.6b-base # Download with progress izwi pull qwen3-asr-0.6b

Via Web UI

  1. Open http://localhost:8080
  2. Go to Models in the sidebar
  3. Click Download on any model

Managing Models

View Downloaded Models

izwi list --local

Get Model Information

izwi models info qwen3-tts-0.6b-base

Load a Model into Memory

izwi models load qwen3-tts-0.6b-base

Unload a Model

izwi models unload qwen3-tts-0.6b-base

Delete a Model

izwi rm qwen3-tts-0.6b-base

Model Storage

Models are stored in your system's application data directory:

PlatformLocation
macOS~/Library/Application Support/izwi/models/
Linux~/.local/share/izwi/models/
Windows%APPDATA%\izwi\models\

Custom Model Directory

Set a custom location:

# Via CLI flag izwi serve --models-dir /path/to/models # Via environment variable export IZWI_MODELS_DIR=/path/to/models izwi serve

Manual Downloads

Some models require manual download from Hugging Face due to licensing:

Model Status

Models can be in several states:

StatusDescription
not_downloadedModel available but not on disk
downloadingCurrently downloading
downloadedOn disk but not loaded
loadingBeing loaded into memory
readyLoaded and ready for inference

Check status:

izwi status --detailed

Quantized Models

Some models offer quantized variants for reduced memory usage:

  • 4-bit — Smallest size, some quality loss
  • 8-bit — Balanced size and quality
  • Full — Original quality, largest size

Quantized models have suffixes like -4bit or -q4.

Next Steps