https://video.ut0pia.org/videos/watch/e781dc55-8c81-45da-b54a-4f0c8ef01d0f Qwen3-TTS shipped at 0.8x real time: one second of audio took 1.2 seconds to generate. Andres Marafioti from Hugging Face spent two weeks fixing it. The culprits were no streaming, 500 autoregressive steps per audio packet with a CPU GPU round trip on each, and a dynamic KV cache that blocked compilation. Static KV cache plus CUDA graph captures brought it to 5.8x real time with time to first audio under 200 milliseconds. The platform is Reachy Mini, a $300 open source robot Hugging Face has shipped to 7,500 people. It arrives unassembled. Talking to it is their most used app by far. The voice stack runs Parakeet transcription every 150 milliseconds with partial results feeding back to the robot mid-sentence, Qwen 3.5 27B for the LLM, and this optimized TTS. At that speed, infrastructure round trips match model latency, so the load balancer separates LLM endpoints from conversation nodes to handle the difference in how much different users actually talk. Sat, 30 May 2026 14:31:54 GMT https://validator.w3.org/feed/docs/rss2.html PeerTube - https://video.ut0pia.org https://video.ut0pia.org/lazy-static/avatars/0287a09a-aae7-4840-9843-b416426e7046.webp https://video.ut0pia.org/videos/watch/e781dc55-8c81-45da-b54a-4f0c8ef01d0f All rights reserved, unless otherwise specified in the terms specified at https://video.ut0pia.org/about and potential licenses granted by each content's rightholder.