KittenML has released Kitten-TTS 2 as open source, a text-to-speech model built around 1.58-bit ternary quantization and the first TTS model to use this quantization level. It runs in real time on a CPU, despite having 1.7 billion parameters. The practical consequence is that self-hosting becomes realistic without a GPU.

Technically, every linear weight in the model body takes one of three values (-1, 0, or +1), with a group size of 128 for scales. Optimized ternary kernels for CPU inference are included. The performance bottleneck shifts from VRAM availability to the quality of the CPU execution path, which changes how teams think about hardware for voice services.

The model supports zero-shot voice cloning from less than five seconds of reference audio, retaining expressiveness and accent without fine-tuning. It covers more than 20 languages and allows cross-language cloning to produce accented speech. Delivery and emphasis control lets users stress specific words, not just pick a voice.

For on-premise deployment, the shift is clear: Kitten-TTS 2 moves local speech generation from GPU to CPU. Teams do not need cloud instances or expensive accelerator cards; a CPU-based architecture can generate audio in real time. This changes the TCO calculation for internal voice assistants, voicebots, and accessibility tools. Voice data can remain inside the organization's infrastructure, which matters for data sovereignty and GDPR compliance.

The technical signal goes beyond the product. Using ternary quantization on a 1.7 billion parameter TTS model shows that aggressive compression is becoming a practical architectural choice rather than a research topic. If the optimized CPU kernels deliver the promised responsiveness, the main constraint for local speech synthesis stops being raw compute and becomes software integration quality. Cloud voice cloning vendors may face pressure on commoditized multilingual use cases, while enterprises with sensitive voice data gain a credible self-hosted path.

KittenML presents the project as an open-source alternative to ElevenLabs for the GPU-poor. The key question is not only whether it matches closed commercial systems on every language or accent, but whether it makes a self-hosted pipeline viable without dedicated hardware. For organizations working with sensitive voice data, keeping the entire pipeline local, from reference audio to generation, changes the trade-off between cloud APIs and infrastructure under direct control. If real-time latency holds up on consumer CPUs under sustained load, this release could set a precedent for local generative audio models.