100B total, 6B active per token in the current model card; the technical report describes 100B / 6.1B, while the launch blog describes 103B / 9B
32,768 tokens
Hugging Face metadata identifies MIT, but the repository has no root LICENSE file and the complete incorporated-component licensing was not independently resolved.
A sparse any-to-any multimodal preview built on the Ling-Flash-2.0 MoE architecture. It accepts text, images, video and audio and generates text, speech and images. The release adds streaming video conversation, contextual and dialect-aware speech recognition, voice cloning, high-fidelity text rendering in generated images, stronger scene and identity consistency during image editing, and generative segmentation-as-editing. The publisher reports improvements across all twelve ContextASR benchmarks and recognition support for fifteen Chinese dialects. Its shipped language-model configuration has 32 transformer layers, 256 routed experts with eight selected per token plus one shared expert, grouped-query attention, and 32,768 maximum position embeddings. Forty-two root BF16 weight shards and the associated connector, speech and diffusion assets are publicly downloadable from Hugging Face and ModelScope. The later Ming-flash-omni-2.0 release is a separate successor and is not grouped into this card.