Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU
First reported by The Decoder ·
An AI model capable of performing at the level of top Western benchmarks is now available for commercial use without requiring foreign hardware. The training of this model on a cluster of 194 Huawei Ascend accelerators, with no guidance for Nvidia GPUs, signals a significant advancement in China's domestic AI hardware and software ecosystem.
Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter AI model designed for autonomous agents, particularly in coding tasks. Notably, the model was trained entirely on Chinese hardware, specifically Huawei Ascend accelerators, and did not use any Nvidia GPUs. LongCat-2.5 is a mixture-of-experts model, activating a fraction of its parameters per token, and supports processing up to one million tokens with an output length of 128,000 tokens. It possesses multimodal capabilities, understanding text, images, and video, and its weights are available under the MIT license for commercial use. Benchmarks show it performing comparably to Western models like Claude Opus 5 on several coding and agent tasks, though it slightly trails on some complex command-line operations and bug-fixing tests. Meituan offers API access with pricing structured per million tokens.
The development and release of LongCat-2.5, particularly its training on non-Nvidia hardware, underscore a growing trend of AI model development uncoupled from traditional Western infrastructure. This could lead to increased competition and diversification in the AI hardware market, potentially lowering costs and increasing accessibility for organizations that previously faced geopolitical or economic barriers to high-performance AI. The model's success on benchmarks comparable to leading US models, despite its unique training origin, challenges the notion that cutting-edge AI development is exclusively tied to US-based hardware and research.
Furthermore, Meituan's approach to multimodal integration, where images are converted into tokens processed alongside text, rather than bolted on via separate encoders, represents a more unified architecture that could improve efficiency and performance in visual understanding tasks. The model's ability to handle very long contexts (up to one million tokens) and its deliberate training methodologies, which include blocking internet access and simulating agent failures, suggest a focus on robust and reliable agentic behavior. This positions LongCat-2.5 as a strong contender for applications requiring sophisticated, multi-step reasoning and task completion in complex environments.
AI-written summary. May contain errors.