LongCat-2.0: China’s Meituan Trains a 1.6-Trillion-Parameter AI on Domestic Chips

by Zaira Varghese 01. July 2026 AI 0

China just cleared the exact milestone US export controls were designed to prevent. On June 30, 2026, super-app giant Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter large language model the company says was trained from start to finish on domestic chips — with no Nvidia hardware in the loop. It is the boldest signal yet that China’s AI industry can build frontier-scale models without Western silicon.

What LongCat-2.0 actually is

LongCat-2.0 is a 1.6-trillion-parameter model with a one-million-token context window, putting it on par with DeepSeek’s current flagship, V4-pro. Meituan trained it on a 50,000-card cluster of domestic AI ASIC “superpods” and, crucially, released the model weights publicly so anyone can download, run, and fine-tune it. The company is pitching it as a coding-and-reasoning workhorse rather than a consumer chatbot.

Parameters: 1.6 trillion (mixture-of-experts architecture)
Context window: up to 1 million tokens
Training hardware: ~50,000 domestic AI accelerator cards
License: open weights, freely downloadable

Why training on domestic chips matters

The headline isn’t the parameter count — it’s the silicon. Meituan’s stack reportedly leans on Huawei technology, including the Huawei Collective Communication Library and Atlas-950 SuperPods. That makes LongCat-2.0 the first trillion-parameter model claimed to complete both full pre-training and inference on Chinese hardware. That distinction matters: DeepSeek-V4-pro used home-grown chips only for inference, while pre-training a frontier model is far more compute-intensive and has historically demanded high-end Nvidia GPUs now blocked from export to China.

The DeepSeek playbook, repeated

By open-sourcing the weights, Meituan is running the same play that turned DeepSeek into a global story earlier this year: give the model away, seed adoption everywhere, and let developer momentum do the marketing. Open weights also make the domestic-chip claim harder to dismiss, because outside researchers can benchmark the model directly. According to SiliconANGLE’s report on the launch, the release lands squarely in the middle of an intensifying US-China race over who controls the compute behind advanced AI.

What it means for the chip war

For Washington, LongCat-2.0 is an uncomfortable data point. The core theory behind export controls was that denying China top Nvidia GPUs would keep it a step behind on training. A trillion-parameter model trained end-to-end on domestic accelerators suggests that gap is narrowing faster than expected — even if efficiency, yield, and real-world performance still lag the best Western clusters. For Nvidia, it is another reminder that its most restricted market is busy engineering around it.

The takeaway

Whether LongCat-2.0 truly matches Western frontier models will take independent testing to confirm, and “trained on domestic chips” leaves plenty of room for asterisks. But the symbolism is hard to overstate: China now has a homegrown, open-weight, trillion-parameter model and the supply chain to train more. The AI race just became a lot less dependent on Nvidia’s export queue.