Open Models Almanac
Methodspost-trainingmid-training & continual pretraining

single mention · post-training

Multimodal mid-training

Multimodal mid-training is a training stage after language mid-training that jointly optimizes the language backbone, connector, and vision encoder.1 It reuses the language mid-training data mixture augmented with multimodal data, aiming to fuse modalities and add visual knowledge while preserving language performance.2

How this definition was written

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

  1. we introduce a multimodal mid-training stage that jointly optimizes the language backbone, pre-trained connector, and vision encoder — LFM2 Technical Report
  2. The language mid-training text mixture is reused and augmented with multimodal data — LFM2 Technical Report

Also called joint multimodal mid-training.

Filed under post-training › mid-training & continual pretraining