Izwi v0.1.0-beta: Kicking Off Our Beta Cycle

We've been hard at work refining Izwi, and we're excited to share that we've officially kicked off our v0.1.0-beta cycle. After several alpha releases focused on core functionality, we're now entering the beta phase with a focus on stability, performance, and preparing for a stable release.

What's New in Beta

We've already shipped two beta releases with some significant improvements:

v0.1.0-beta-1

• Flash Attention 2 support for faster inference
• GGUF chat model support for flexible model deployment
• Local/edge scheduler performance improvements with deadline-aware preemption
• Token-level decode microbatching for better throughput
• Qwen3-TTS orchestration with unified history persistence

v0.1.0-beta-2

• Kokoro-82M TTS - A new high-quality text-to-speech model
• Streaming TTS with adaptive chunking for real-time audio
• Qwen3 4B/8B/14B GGUF variants for more deployment options
• WebSocket streaming for voice with binary audio, server VAD, and barge-in
• Complete UI overhaul - New design system, streamlined model management, polished chat/voice history

What's Coming Next

The beta cycle is all about stability and polish. Over the next few weeks, we'll be:

• Squashing bugs and improving reliability
• Performance tuning across all supported hardware
• Expanding model support
• Polishing the desktop and server experience

Get Involved

This is the perfect time to try out Izwi and share your feedback. As an open-source project, your input shapes the future of Izwi.

Join us on this journey to make on-device voice AI accessible to everyone.