Tech Jun 14, 2026 13 min ZONOS2 on an 8GB RTX 4060 Laptop (WSL2): it runs, but ~20x slower than realtime Tested ZONOS2 on an 8GB RTX 4060 Laptop (WSL2): the 15.3GB bf16 weights run via Windows system-memory fallback, a KV-cache override, and the CUDA toolkit at ~1/20 realtime. Plus a Japanese name-accent gotcha with A/B audio. AI TTS Speech Synthesis ZONOS2 Zyphra HuggingFace Japanese
