@juntao
Python is for AI research. Rust is for AI production. We created a complete Rust impl of tiny TTS models (15M-80M) from @divamgupta 10 MB binary, 100ms cold start, CPU-only, edge-ready. π±π¦ https://t.co/7HngBVd0Jw CPU runs KittenTTS at 4x real-time, but Apple's GPU only hits 3x. @claudeai told us why: at this model size, GPU dispatch overhead exceeds the compute savings. ONNX's CPU backend with SIMD already saturates the hardware. Below ~1B params, the GPU loses. Full story π