single mention · model architecture
Temporal-Modality Rotary Position Embedding
TM-RoPE (time-aligned multimodal RoPE, from Qwen2.5-Omni) splits rotary position embedding into time, height and width and ties the time id to the clock, so an audio token and a video frame from the same moment share a temporal position. One id meant 40 ms in Qwen2.5-Omni; the Qwen3.5-Omni report gives 160 ms.
Written for the explainer Hearing, speaking and keeping time: audio, omni models and multimodal positions.
Also called TM-RoPE.
Filed under model architecture › positional encoding