Kitten-TTS 2 released: first 1.58-bit TTS model, real-time on CPU with voice cloning
ElectricalBar7464 · reddit · 2026-10-22
KittenML has open-sourced Kitten-TTS 2, billed as the world's first 1.58-bit (ternary) TTS model — an open ElevenLabs alternative for GPU-poor users.
- Ternary weights: every linear weight is one of three values (-1/0/+1) with group size 128 for scales; 1.7B parameters, tiny footprint
- Real-time on CPU: no GPU needed, with optimized ternary CPU kernels; consumer GPUs also supported
- Zero-shot voice cloning from under 5 seconds of reference audio, preserving expressivity and accent, no fine-tuning
- Multilingual: 20+ languages, with cross-language cloning for accented speech
- Emphasis control: steer delivery and stress specific words
Version 2 uses a completely new architecture with more natural speech over V0.8. Code and weights are fully open for local or self-hosted deployment.
More from Infra
- SF Systems Meetup: Agents-Only Version Control and Postgres Rewritten in Rust — ShadajL · 2026-10-22
- Cloudflare saw thousands of php2shell exploit attempts before disclosure, millions now — ivan_bezdomny · 2026-10-22
- Jensen Huang: NVIDIA Is "Literally the Only One" You Can Trust With a $100B Blind Chip Order — jfischoff · 2026-10-22
- Dual RTX PRO 6000 Benchmarks: DeepSeek-V4-Flash vs Qwen3.8-27B Local Serving — dee_hw · 2026-10-22
- Korean neocloud Elice now up 68% from IPO-day low — the panic headlines have gone quiet — firstadopter · 2026-10-22
- Berth v0.3.4 adds GPU VFIO passthrough: lend your 4090 to a Windows VM and get it back — Distinct-Pie2389 · 2026-10-22