inclusionAI Debuts 'Any-to-Any' Multimodal MoE Model

AI research group inclusionAI has released Ming-flash-omni-Preview, a new open-source model designed for true multimodal flexibility. Released under a permissive MIT license, the model pursues an "any-to-any" capability, meaning it's built to process and generate a wide combination of data types, not just text and images.

This approach, often called "omnimodal," represents a significant step beyond models that are limited to specific input-output pairs, like text-to-image or audio-to-text. An any-to-any system can theoretically accept a mix of inputs—say, an image, a line of text, and an audio clip—and generate a relevant output in a requested modality.

The model is built on a Mixture of Experts (MoE) architecture, a technique that improves computational efficiency by routing inputs to specialized subnetworks, or "experts," rather than engaging the entire model for every token. According to the release card, Ming-flash-omni is based on a previous model called Ling-flash-2.0.

As major labs pursue closed, highly capable omnimodal models, the release of an open alternative like Ming-flash-omni-Preview provides researchers and developers with a valuable tool for experimentation. While labeled as a preview, it offers a foundational component for building applications that require a more fluid and comprehensive understanding of diverse data streams.