20.3B total, 3B active in the MoE section
32,768 tokens
Hugging Face metadata identifies MIT, but the repository has no root LICENSE file. The packaged talker includes Apache-2.0 terms, while the predecessor's bundled connector carried the non-commercial Qwen Research License; applicability of that earlier component license to the v1.5 package remains unresolved.
An any-to-any multimodal model built on Ling-lite-1.5 that accepts text, images, video and audio and can generate text, speech and images. Version 1.5 upgrades the earlier Ming-Lite-Omni with MRoPE-based spatiotemporal encoding and curriculum learning for long-video understanding; dual-branch image generation and editing with identity- and scene-consistency objectives; a new audio decoder and BPE-based speech representation for real-time speech synthesis; and expanded structured-text, product, visual and speech training data. The official model card reports a 20.3B-parameter system with 3B active parameters in its MoE section and reports a 69.83 average on VideoMME, alongside improvements in document understanding, OCR, chart reasoning and multimodal generation. The shipped language-model configuration uses 32,768 maximum position embeddings. Downloadable BF16 weights are provided through both Hugging Face and ModelScope, with a separately released FP8 package excluded from the headline parameter count.