🆕vLLM-Omni! Your RTX 5090 could give your Agent a quicker voice. One developer measured 47 ms to first audio. A community developer reports the median result with Qwen3-TTS 0.6B-CustomVoice. Audio starts streaming while the model is still generating later chunks. Your app can begin playback before the whole sentence is ready. For Local AI users building voice assistants, that’s a practical piece of the puzzle: getting speech out quickly on hardware you own. vLLM’s new Omni technical report explains the machinery for coordinating text, audio, image and video generation. The demo includes code, audio samples and installation instructions. Its setup requires Linux and at least 10GB of free VRAM. Link in ALT.