Meituan open-sources the omni-modal model LongCat-Flash-Omni
meituan-longcat/LongCat-Flash-Omni-FP8
AI BriefingMeituan open-sourced the fully multimodal model LongCat-Flash-Omni, using a 560-billion-parameter MoE architecture (27 billion activated parameters), with support for a 128K context and real-time audio and video interaction. The model uses shortcut-connected MoE and zero-computation experts, equipped with a lightweight codec and a chunked feature interleaving mechanism, and improves efficiency through curriculum-based progressive training. It outperforms Qwen3-Omni and Gemini-2.5-Pro on benchmarks including OmniBench and WorldSense, reaching leading performance in tasks such as document understanding, speech recognition, and GUI control.
Source: Meituan LongCat: New HuggingFace model · huggingface.co