Meituan

TensorWorld, founded just four months ago, has raised over $100 million from investors including Meituan, Ant Group and NIO Capital. It pursues a physics-driven world-model approach that learns physical dynamics from data rather than chasing visual realism, aiming to support robot action planning and long-horizon prediction. However, key technical details are largely omitted and the core method remains unverified publicly.

Meituan LongCat has open-sourced LongCat-2.5, a natively multimodal model with 1.6 trillion total parameters, roughly 48 billion activated per token, and one-million-token context, released under the MIT license for commercial inference and retraining. It encodes visual signals as multi-level discrete tokens sharing the same backbone as text, and uses a W4A4 4-bit quantization scheme tuned for domestic AI chips constrained by memory and bandwidth.