Tiny AI model creates human voice clones from a web browser.

AI company Zyphra has released ZONOS3-Mini, an open-source text-to-speech model designed for local, real-time use. This model uses a 'Mixture of Experts' architecture, boasting 800 million total parameters but only activating about 100 million per task, which makes it surprisingly lightweight. It supports zero-shot voice cloning—meaning it can mimic a new voice from a short sample—and allows precise control over word-level delivery. The model's weights are released under the permissive Apache 2.0 license, inviting widespread developer use.

For engineers, the breakthrough is efficiency: this model runs directly in a web browser via WebGPU, on laptops with Apple Silicon using MLX, or on servers with NVIDIA GPUs. This opens the door for high-quality, responsive TTS in applications without relying on cloud API calls, from interactive games to assistive tech. However, all performance benchmarks are self-reported by Zyphra, and its reference-clone scoring lags behind other modes. The takeaway is a significant step toward democratizing professional-grade, responsive speech synthesis that runs directly on user devices.