Get App
Download App Scanner
Scan to Download
Advertisement

China's Meituan Trains 'Country's Largest' 1.6-Trillion-Parameter AI Model Without Nvidia Chips

Meituan said LongCat-2.0 outperformed Google's older Gemini 3.1 Pro on several coding and agentic benchmarks, though it trailed global frontier models including OpenAI's GPT-5.5 and Anthropic's Claude 4.8 Opus.

China's Meituan Trains 'Country's Largest' 1.6-Trillion-Parameter AI Model Without Nvidia Chips
Meituan said LongCat-2.0 outperformed Google's older Gemini 3.1 Pro on several coding benchmarks.
Igor Omilaev/ Unsplash

Chinese food delivery giant Meituan has built and trained a 1.6-trillion-parameter artificial intelligence model entirely on domestic chips, bypassing Nvidia hardware altogether.

Meituan claims that it is the country's largest AI system developed without Western processors, the South China Morning Post reported on Tuesday.

The Beijing-based company open-sourced the model, called LongCat-2.0, boasting 1.6 trillion parameters and a context window of one million tokens, putting it on par with DeepSeek's flagship V4-pro model launched in April, according to SCMP.

Full Training Cycle On Chinese Hardware

Meituan said LongCat-2.0 was the industry's first trillion-parameter model to complete both pre-training and inference entirely on a 50,000-card domestic computing cluster, the report said.

While DeepSeek-V4-pro used home-grown chips only for inference, Meituan's model reportedly relied on Chinese hardware for pre-training as well, a far more computationally demanding process in which a model learns from massive datasets. 

The company said the model was built on large-scale clusters of AI application-specific integrated circuits (ASICs), chips customised for particular workloads rather than general-purpose processors.

Though Meituan did not name its supplier, it disclosed using Huawei's Collective Communication Library to improve training stability, according to the report.

Meituan said LongCat-2.0 outperformed Google's older Gemini 3.1 Pro on several coding and agentic benchmarks, though it trailed global frontier models including OpenAI's GPT-5.5 and Anthropic's Claude 4.8 Opus.

ALSO READ: US Lifts Export Controls On Anthropic's Claude Fable 5 And Mythos 5

Steep Hurdles Remain

Despite the milestone, Meituan acknowledged persistent challenges in moving away from Nvidia's ecosystem, noting that its domestic accelerators carried significantly less memory per device than Nvidia's export-banned H800 chip. 

The company said memory was the "primary bottleneck" in training on its 50,000-chip cluster, requiring extensive optimisation to build stable infrastructure.

The breakthrough comes as Beijing pushes for technology self-reliance amid continued US export restrictions on advanced semiconductors to Chinese firms.

ALSO READ: Jana Holdings' NCD Repayment Delay Unlikely To Affect Jana SFB, Sources Say

Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.

Newsletters

Update Email
to get newsletters straight to your inbox
⚠️ Add your Email ID to receive Newsletters
Note: You will be signed up automatically after adding email

News for You

Set as Trusted Source
on Google Search
Add NDTV Profit As Google Preferred Source