While the industry is still debating whether multimodal AI can truly take off, Meituan has quietly unveiled a strong card - the newly open-sourced large model LongCat-Flash-Omni is officially launched and has surpassed several closed-source competitors in multiple benchmark tests, achieving a rare breakthrough of "open source as SOTA" (State-of-the-Art). This AI system, whose name implies "versatility," not only supports real-time integration of text, speech, images, and videos, but also pushes local multimodal intelligence to a new height with near-zero latency interaction experience.

The impressive aspect of LongCat-Flash-Omni lies in its precise control over complex cross-modal tasks. Test results show that when facing questions such as "describing the motion trajectory of a small ball within a hexagonal space," which combines physical logic and spatial reasoning, the model can not only accurately model it but also clearly explain the dynamics process in natural language. In speech recognition, even in high-noise environments, it can accurately extract semantics; when dealing with blurry images or short video clips, it can quickly locate key information and generate structured answers.

image.png

All of this is due to its innovative end-to-end unified architecture. Unlike traditional multimodal models that process each modality branch independently and then concatenate them, LongCat adopts an integrated design, allowing text, audio, and visual data to align and reason within a unified representation space. During training, the team used a progressive multimodal injection strategy - first solidifying the language foundation, then gradually introducing image, speech, and video data, enabling the model to maintain language capabilities while steadily improving cross-modal generalization performance.

What's more surprising is its extreme optimization of response speed. Thanks to the Flash inference engine and lightweight design, LongCat-Flash-Omni can achieve smooth conversation on ordinary consumer-grade GPUs. When users experience it through the official LongCat app or web version on Meituan, they almost feel no delay between input and response, truly achieving the natural interaction of "what you ask is what you get."

image.png

Currently, the model is freely available on Meituan's platforms, developers can obtain the weights through Hugging Face, and ordinary users can directly try it within the application. This move not only demonstrates Meituan's confidence in AI infrastructure technology, but also signals a clear intention to promote the development of the domestic multimodal ecosystem.

At this critical moment when AI competition is shifting from "single-modal accuracy" to "multimodal collaboration," the emergence of LongCat-Flash-Omni represents both a breakthrough in technical boundaries and a redefinition of application scenarios. When a food delivery platform can train a multimodal large model that rivals international giants, the second half of China's AI journey may have just begun.