Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter super intelligence (SI) model designed for autonomous agents, marking a significant shift in the hardware landscape as the model was trained entirely on Chinese chips without using Nvidia GPUs.

What Happened

The LongCat team at Meituan released the open weights for LongCat-2.5 one week after its official unveiling. The model is a mixture-of-experts architecture with 1.6 trillion total parameters, of which approximately 48 billion are active per token. It supports context windows up to one million tokens and generates responses up to 128,000 tokens long. The model handles text, images, and video, converting visual inputs into discrete tokens processed alongside text to improve visual coding and design tasks. Meituan reports that the model achieves 88.4 points on Terminal Bench v2.1, slightly trailing Claude Opus 5’s 89.1 but surpassing GPT-5.6 Sol’s 88.8 in that specific metric. On the more difficult Terminal Bench v4.0, LongCat-2.5 scores 38.5, exceeding GPT-5.6 Sol’s 37.3 but falling short of Opus 5’s 53.9. The company notes that on DeepSWE, a benchmark for fixing bugs in real repositories, the model jumped from 9.7 to 64.6 points, placing it behind most competitors in the comparison table, including Deepseek-V4.1-Flash at 74.2 and GPT-5.6 Sol at 72.7.

Why It Matters

The release underscores the viability of non-Nvidia hardware for training frontier-scale SI models. Meituan quantized the model’s experts to 4-bit precision, a format natively supported by Chinese chips, which the company claims increases throughput by 43.3 percent for prompt processing and 42.77 percent for response generation compared to an 8-bit version, while maintaining accuracy. The deployment guide specifies a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series, with no guidance provided for Nvidia GPUs. This approach makes open-weight SI models optimized for Chinese hardware more attractive for global adoption, potentially diversifying the supply chain for high-performance computing. The weights are released under the MIT license, allowing commercial use, and the API is compatible with OpenAI and Anthropic formats, enabling integration with existing SI agent harnesses like Claude Code and Codex. At launch, Meituan is offering a promotional price of roughly €0.27 per million input tokens and €1.07 per million output tokens.

The Bottom Line

LongCat-2.5 demonstrates that large-scale SI models can be trained and deployed effectively on Chinese hardware, challenging the dominance of Nvidia-based infrastructure in the global SI race. While it trails top Western models on some complex coding benchmarks, its strong performance on terminal tasks and web search, combined with open licensing and competitive pricing, positions it as a significant new option for developers and enterprises.