Meituan’s LongCat-2.0: China’s 1.6T AI Runs on Native Chips

Imagine building a colossal language model without a single Nvidia card in the rack. Strange? Not anymore. Meituan quietly dropped LongCat-2.0, an open-source giant that challenges assumptions about which hardware can shoulder the largest AI workloads.

LongCat-2.0 packs roughly 1.6 trillion parameters and supports a context window of about one million tokens, placing it squarely among the most expansive language models in circulation. In scale, its creators say it rivals models like DeepSeek-V4-pro, but the headline isn’t raw size. It’s the hardware story.

Unlike prior efforts that leaned on domestic chips only for inference, Meituan reports that LongCat-2.0 completed both training and inference on an entirely China-based stack. The company claims to have used a compute cluster built from some 50,000 Chinese graphics cards and tens of thousands of ASICs tailored for AI tasks. The result: large-scale training carried out on hardware alternatives to the global Nvidia-dominated supply chain.

The firm has not disclosed the exact processor families involved. Huawei is an obvious candidate, but Meituan stopped short of naming suppliers. That opacity matters: software compatibility, driver maturity and orchestration layers all shape real-world performance, not just the chips themselves.

LongCat-2.0 marks a milestone in hardware independence, but its practical capabilities remain unproven until independent benchmarks arrive.

So far the model hasn’t been run through leading evaluation suites such as Artificial Analysis or Arena, nor through specialized tests like Agents' Last Exam and CyberGym. Benchmarks will tell whether training on a native stack translates into competitive accuracy, latency and efficiency. Will this be parity with Nvidia-trained counterparts, or a different trade-off entirely?

The broader implication is clear: China’s ecosystem is testing whether alternative chips and massive domestic clusters can sustain frontier AI research. That could reshape supply-chain assumptions and spur local innovation. For now, LongCat-2.0 is a proof point of capability and a prompt to watch the benchmark results that will decide whether it is a breakthrough or an intriguing experiment.