Meituan's 1.6-trillion-parameter AI model runs entirely on Chinese hardware
Meituan launched LongCat-2.5 on November 15, 2026—a 1.6-trillion-parameter agent model trained using over 50,000 Chinese AI chips. It operates without any Nvidia GPUs, processing up to one million tokens per API call and generating responses up to 128,000 tokens long. The model quantizes weights and activations to 4-bit precision for native execution on Huawei Ascend Atlas A2 accelerators (194 total: 2 for images, 64 for prompts, 128 for responses). It converts images directly into tokens for multimodal processing, unlike most models. LongCat-2.5 runs under MIT license and offers promotional pricing of roughly €0.27 per million input tokens and €1.07 per million output tokens. Benchmarks show it achieves 88.4 on Terminal Bench v2.1 (vs. Claude Opus 5's 89.1) and 91.9 on Meituan's internal BrowseComp test.
This model directly reduces China's reliance on foreign AI hardware. By training exclusively on domestic infrastructure, it cuts costs for Chinese AI services while enabling faster deployment of local applications. The pricing structure—though promotional—suggests potential cost advantages for businesses using the model within China's ecosystem.
What to watch: Meituan's accuracy claims after 4-bit quantization (caveats note it remains comparable to 8-bit versions) and how Terminal Bench v4.0 scores (a harder variant) impact real-world utility. The model's multimodal capabilities could accelerate domestic AI adoption, but its performance gaps against newer Western models like GPT-6.1 Sol mean broader adoption depends on practical use cases. The source specifies pricing is approximate, so exact cost impacts remain unverified.
Source: The Decoder
Many Minded