Meituan claims its LongCat-2.0 is the first trillion-parameter AI model trained entirely on domestic chips, sidestepping US export controls that have choked access to NVIDIA's best hardware.
A trillion-parameter milestone on homegrown silicon
LongCat-2.0 isn't just another large language model. At a trillion parameters, it lands in the same weight class as Google's Gemini 3.1 Pro, a frontier model that typically requires high-bandwidth memory and cutting-edge accelerators. Meituan says it achieved this using only domestically developed processors — no names given, but likely from Huawei or another sanctioned Chinese fab. That's notable because training a trillion-parameter model from scratch usually demands clusters of H100s or B200s, exactly the chips the US Commerce Department has restricted since October 2022.
Why this cuts against the conventional wisdom
Most analysts assumed China would hit a ceiling just below the trillion-parameter mark without access to advanced lithography or NVIDIA's software stack. Meituan just shoved that assumption in a drawer. If LongCat-2.0 really delivers capability comparable to Gemini 3.1 Pro on inference benchmarks — and the company hasn't published numbers yet — then the export controls have a thinner margin than many believed. The trick likely involves aggressive model parallelism, clever memory management, and a willingness to train for weeks longer on slower interconnects. Whatever the engineering tradeoffs, the result is a functional trillion-parameter model that runs on Chinese silicon.
What this unlocks for Meituan and China's AI supply chain
Meituan isn't a research lab; it's a food delivery and services giant with hundreds of millions of users. Shipping a model this size means they intend to deploy it at scale, likely for recommendation systems, logistics optimization, and conversational AI. If LongCat-2.0 goes into production, it validates that domestic chip ecosystems can sustain frontier-level AI workloads. Other Chinese firms — ByteDance, Alibaba, Baidu — are watching closely. The next six months will show whether this is a one-off engineering feat or the start of a genuine alternative supply chain.
Closing the gap on hardware doesn't mean China has caught up on software tooling or ease of development. But LongCat-2.0 proves the ceiling is higher than the export controls were designed to impose.
Source: China's Meituan says new AI model trained on domestic chips
Domain: economictimes.indiatimes.com
Comments load interactively on the live page.