inclusionAI Releases Ming-UniVision MoE Multimodal Model

inclusionAI has released Ming-UniVision-16B-A3B, a new open-source model designed for flexible multimodal tasks. Released under the permissive Apache 2.0 license, it introduces an efficient architecture for handling diverse data types beyond simple text and image pairings.

The model employs a Mixture-of-Experts (MoE) design, with a total of 16 billion parameters. During inference, however, it only activates a fraction of those—just 3 billion, a detail likely referenced by the "A3B" in its name. This sparse activation approach aims to provide the power of a very large model while keeping computational costs relatively low.

What sets Ming-UniVision apart is its goal of being an "any-to-any" model. Unlike many vision-language models that are limited to specific input-output paths, like image-to-text, this architecture is built to process and generate a wider combination of data formats. This makes it a more versatile tool for complex, multi-faceted AI tasks.

Availability

The model and its weights are available for download on Hugging Face. Its open license encourages broad adoption for both academic research and commercial applications, paving the way for more sophisticated and efficient multimodal systems.