vLLM-Omni is the vLLM project’s open-source framework for efficient inference with omni-modality models — systems that understand and generate text, images, audio and video in one model. It has around 7,300 stars on GitHub and is under heavy active development in 2026.
The original vLLM became the standard high-throughput serving engine for text LLMs, but a new generation of omni models — which chat, see, listen and generate media — does not fit a text-only serving stack. vLLM-Omni extends the vLLM architecture to these multimodal workloads, providing a unified framework for serving models that combine language modeling with diffusion-based image and video generation and audio generation. It brings vLLM’s signature efficiency techniques — paged attention, continuous batching and optimized kernels — to heterogeneous pipelines where an autoregressive model and a diffusion model must work together in a single request flow. The project ships documentation, examples and benchmarks, and evolves quickly alongside the vLLM ecosystem. It is free and open source under the Apache-2.0 license.
vLLM-Omni is completely free and open source under the Apache-2.0 license. You run it on your own GPU infrastructure; there is no hosted tier or license fee.
vLLM-Omni is for ML platform engineers and researchers who need to deploy multimodal models in production and want vLLM-class throughput and efficiency for workloads that go beyond plain text generation.