Can a single model generate all modalities without external encoders?

Most multimodal LLMs are dual-modal (text + one other modality) and rely on external encoders with alignment modules — they understand non-text inputs but rarely generate them, and can't produce interleaved multimodal sequences. MIO addresses this by training on a mixture of discrete tokens across four modalities (text, image, video, speech) with causal multimodal modeling, through a four-stage process (alignment → interleaved → speech-enhanced pretraining → comprehensive SFT). The result is end-to-end, autoregressive, any-to-any understanding and generation in one model, with emergent abilities the dual-modal baselines lack: interleaved video-text generation and chain-of-visual-thought reasoning.

The keeper is the design point: tokenize every modality into a shared discrete vocabulary and model them with one causal objective — which unlocks interleaved cross-modal output and reasoning that staple-on encoder approaches can't produce.

This sits in the vault's multimodal-architecture thread as the discrete-token, any-to-any design point. It is the tokenizer-based ("Type-D") instantiation contrasted with encoder-based unified generation, and it complements Why do unified image generators fail on non-Latin scripts?: unified any-to-any generation still inherits the frequency-driven data-bias documented in Does multimodal zero-shot performance actually generalize or interpolate?.