Welcome to vLLM office hours! These bi-weekly sessions are your chance to stay current with the vLLM ecosystem, ask questions, and hear directly from contributors and power users. This week's special topic: vLLM Omni Project Update & Demos. vLLM project update from core maintainer Michael Goin, covering a recap of the vLLM Conference, recent work from the SkyRL team on reinforcement learning and large scale weight transfer, the updated AgentX benchmark, and day-zero support for a wave of new model releases. Plus the v0.28 release itself, with a major Kimi K3 performance push, decode context parallelism, speculative decoding advances, and continued Model Runner V2 maturation. Lightning talk: 8K EDU, a winning project from the recent vLLM hackathon in Austin, which turns frames from any educational video into interactive widgets and notebooks you can actually play with. Then the main event with the vLLM Omni team: what vLLM Omni is and why it exists, how it extends vLLM with a stage-based pipeline for models that take in and produce audio, images, and video, and a Diffusion 101 walkthrough on why diffusion needs its own scheduler and KV cache abstraction. The session closes with two live demos: a real-time voice assistant built on Qwen3-Omni running on a single GPU, comparing the real-time API against chat completions, and a side-by-side video generation demo showing high quality generation with synchronized audio alongside a real-time streaming model. Slides: https://docs.google.com/presentation/d/1L84FWrGAI-fns6TaR0Q8R6I9CvbrWQvbVbGl_KtXhy8/ Want to join the discussion live on Google Meet? Get a calendar invite by filling out this form: https://red.ht/office-hours Timestamps: 00:00 Intro 01:42 vLLM project update 11:27 vLLM v0.28 release 15:50 Lightning talk: 8K EDU hackathon project 24:00 vLLM Omni project update 30:02 Diffusion 101 43:03 Demo: Qwen3-Omni voice assistant 50:00 Demo: video generation side by side

 1091      9