Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU

What changed

Meituan announced LongCat-2.5, a 1.6-trillion-parameter AI agent model trained entirely on domestic Chinese hardware without using Nvidia GPUs. The model’s training leveraged 194 Huawei Ascend chips, demonstrating a large-scale AI effort independent of Nvidia’s established GPU ecosystem. LongCat-2.5 achieves an 88.4 score on Terminal Bench, placing it just behind Claude Opus 5, a notable benchmark for AI agents.

Why builders should care

This development shifts part of the AI model training landscape by showing a major player can build and train extremely large models on non-Nvidia hardware. For developers and operators, it signals that alternatives to Nvidia’s GPU monopoly are emerging, especially in markets prioritizing hardware sovereignty. This could reduce reliance on U.S.-based chip designs, lower geopolitical risk, and open new options for AI workloads in China and elsewhere.

The practical takeaway

Operationally, LongCat-2.5 suggests cost structures and performance trade-offs for massive models might diversify beyond GPU-heavy setups. Organizations building AI agents or large language models should monitor non-GPU hardware advancements, especially in sectors where supply chain risks or hardware localization are strategic concerns. This model also indicates that AI agents approaching top-tier benchmarks no longer require Nvidia GPUs, potentially influencing infrastructure investments.

What to watch next

Keep an eye on performance improvements in AI models running on alternative hardware like Huawei Ascend chips and how the Chinese hardware ecosystem scales for broader AI adoption. Watch how this effort influences Nvidia’s competitive positioning in the global AI infrastructure market and whether similar large models emerge that bypass Nvidia GPUs elsewhere. Also, observe Meituan’s next deployments to see how this model shapes real-world agent applications.

AI Quick Briefs Editorial Desk