IT Home: Moore Threads GPU Adapted to Meituan's LongCat-2.0 AI Model
IT Home (China) reports that Moore Threads' MTT S5000 GPU and MUSA software stack have been adapted for Meituan's newly open-sourced LongCat-2.0 large AI model, enabling efficient inference.
Chinese technology news outlet IT Home reported on July 6 that Moore Threads has completed the adaptation of Meituan's recently open-sourced LongCat-2.0 large artificial intelligence model to its hardware.
The adaptation process covers the entire chain, from model loading and inference engine startup to the optimization of critical computational operations and deployment verification. The LongCat-2.0 model can now run efficiently on Moore Threads' MTT S5000 AI accelerator cards, which are designed for AI training and inference.
LongCat-2.0 is a new-generation MoE (Mixture of Experts) model developed by Meituan, featuring over a trillion parameters. It is specifically designed for agentic coding tasks and natively supports context windows of up to one million tokens, employing mechanisms like sparse attention to enhance resource utilization.
Moore Threads stated that the MTT S5000 cards, with their native FP8 acceleration and large memory capacity, combined with the MUSA software stack, provide a stable foundation for processing long contexts and improving response times. The company highlighted that a standardized adaptation process accelerates the deployment of cutting-edge models on domestic computing platforms.