Open Models Almanac
Methodsinference & serving

single mention · inference & serving

Interleaved text-and-audio generation

Interleaved text-and-audio generation is an inference mode in which the model emits text and audio tokens alternately within one response, following a fixed pattern.1 It works like a chain-of-thought: the model first produces the semantic content in text, then translates that text into audio.2

Trade-off The fixed text-to-audio ratio keeps text generation ahead of audio and reduces time to the first audio token.3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. generates the assistant response in both text and audio, with a predetermined interleaved pattern — LFM2 Technical Report
  2. This acts as a kind of chain-of-thought, where the model generates the semantic content in text — LFM2 Technical Report
  3. The ratio of text and audio tokens is selected to keep text generation ahead of audio generation and to minimize time-to-first-audio token — LFM2 Technical Report

Also called interleaved.

Filed under inference & serving