Qwen2.5-Omni

Alibaba's Qwen team designed Qwen2.5-Omni around a Thinker-Talker split: the Thinker handles understanding across text, image, audio, and video inputs while the Talker generates natural speech in a streaming fashion, with a TMRoPE position embedding aligning video frames to their audio track. The result is a single end-to-end model that can watch, listen, converse, and answer in either text or voice. The 7B model shipped in March 2025 under Apache-2.0, a 3B variant followed in April 2025, and GPTQ-Int4 and AWQ quantizations cut memory use by more than half. Hardware demands are documented precisely: in BF16 the 3B model needs 18.4 GB of VRAM for 15 seconds of video and the 7B model 31.1 GB, rising with clip length, so serious use requires a well-equipped GPU. Deployment paths include HuggingFace transformers, vLLM with audio output support, official Docker images, and MNN for edge devices. With code and weights fully open and no usage fees, it became the leading open any-to-any speech-capable model of its generation and a common base for voice assistant research and multimodal agent prototypes.