Huawei Ascend A2 Runs 1.6T LongCat-2.0 AI Model With 20ms Latency

Huawei has publicly disclosed the technical details of running the LongCat-2.0 large language model on its Ascend A2 hardware, marking a significant step for domestic AI infrastructure. This deployment proves that Chinese-made chips can handle the massive computational load of trillion-parameter models, which matters for enterprises seeking to reduce reliance on foreign silicon. The announcement shifts the conversation from theoretical capability to practical, large-scale inference performance.

Domestic chips handle trillion-parameter model inference

The LongCat-2.0 model, developed by Meituan, features 1.6 trillion parameters and natively supports a context window of 1 million tokens. Huawei’s CANN team deployed this model on a cluster of 192 Ascend Atlas A2 cards, with each server housing 16 cards. This architecture allows the system to manage the complex data flow required for such a large model without collapsing under the weight of its own parameters.

Technical optimizations play a critical role in achieving the system's performance metrics. The CANN team used a combination of pipeline parallelism and sequence parallelism during the prefill stage, splitting the model into eight independent computing stages. They also compressed Mixture-of-Experts (MoE) communication overhead by ten times. During the decode stage, the system differentiated between short and ultra-long requests to maintain efficiency, achieving a time per output token (TPOT) latency of just 20ms.

The Ascend A2 cards provide 200Gbps of network bandwidth per card, enabling the high-speed data transfer necessary for this scale of operation. The deployment demonstrates that domestic chips can handle trillion-parameter large model inference tasks, according to industry insiders. We looked at Apple iPhone 18 Pro Gets A20 earlier while tracking Huawei launches.

Huawei’s disclosure confirms that the Ascend A2 cluster can successfully train and infer the LongCat-2.0 model entirely on domestic computing power. This achievement establishes a benchmark for future large-scale AI deployments within China's domestic hardware ecosystem.