Zyphra has released ZONOS3-Mini, an open-source text-to-speech model utilizing a sparse Mixture of Experts (MoE) architecture with 800 million total parameters and approximately 100 million active parameters per token via 64 experts with top-2 routing. The model supports zero-shot voice cloning, word-level delivery control, and a 32,768-position context window for generations up to 10 minutes. Architectural changes include the Zyphra Audio Codec 1 (ZAC-1) operating at 15 Hz with 10 codebooks (1 semantic, 9 acoustic) and a causal iSTFT head for 44.1 kHz streaming; a hybrid backbone of 11 layers (dimension 1,024) mixing 9 Gated DeltaNet layers with 2 global attention layers plus latent MoE and Q balancing; and the MECT-B2 speaker encoder accepting raw clips up to 60 seconds or 192-dimensional embeddings. Post-training employs DPO followed by GRPO. Deployment targets include WebGPU in browsers, SGLang-Omni on NVIDIA GPUs, MLX on Apple Silicon, and ggml for CPU/CUDA/Metal/Vulkan. Self-reported benchmarks on Zyphra's ZTTS2-Eval show a Capability Index of 66.2 (embedding mode), 7.79 WER, and 83.4 speaker similarity, outperforming Fish Audio S2 Pro (61.7) and MOSS-TTS v1.5 (60.2). Latency reaches 80 ms median time-to-first-audio on one H100 at 8 concurrent requests, versus under 570 ms on a Pixel 9 browser. However, reference-clip cloning scores only 44.7 on the index, and MSR-UTMOS quality (3.07) trails Fish S2 Pro (3.16) and Qwen3-TTS 1.7B (3.54). Weights are Apache 2.0 licensed.