appConfigEntityId

bodyType

customContents

entityId

entityUuid

isVideo

articlePageQuery

article

articleSchemaContent

content

hooksArticlePageAdvertisementConfigArticle

hooksArticlePageFloatingButtonConfigArticle

appConfig

filter

json

liveArticle

Xinmei Shen

As China attempts to move beyond using domestic chips solely for model inference, food delivery giant Meituan released what it claims is the country’s largest artificial intelligence model trained entirely on home-grown hardware.
The Beijing-based on-demand service giant on Tuesday open-sourced LongCat-2.0, a new large language model (LLM) boasting 1.6 trillion parameters and a context window of 1 million tokens. The scale puts it on par with DeepSeek’s latest flagship model, V4-pro, which launched in April.
Meituan claimed that LongCat-2.0 was the industry’s first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster.
While DeepSeek-V4-pro relied on home-grown chips only for inference – the process where a pre-trained model runs to answer user queries – LongCat-2.0 used domestic hardware for both inference and pre-training, according to Meituan.
Pre-training is a far more computationally intensive process, during which an AI model digests massive data sets to learn basic patterns.
Meituan said LongCat-2.0 was built entirely on “large-scale clusters of tens of thousands of AI ASIC superpods”, showing its ability to “conduct frontier-scale training on alternative hardware platforms”. An ASIC, or application-specific integrated circuit, is a chip customised for specific workloads, as opposed to a general-purpose processor.
While Meituan did not explicitly name its hardware supplier, the company said in a WeChat post on Tuesday that it used the Huawei Collective Communication Library (HCCL) to improve training stability. HCCL is a chip-to-chip communication system similar to the Nvidia Collective Communication Library (NCCL).
Meituan and Huawei Technologies did not immediately respond to requests for comment on Tuesday.
China’s home-grown AI chips have been widely adopted for model inference amid Beijing’s push for technology self-reliance. However, local hardware has long been considered inadequate for LLM pre-training, and most Chinese models trained on domestic hardware have been small or limited to multimodal tasks.
Meituan’s breakthrough drew immediate attention from industry observers.
“This put to rest any concerns of Atlas-950 SuperPoDs [being] unable to train large LLMs for [Zhipu AI] and DeepSeek,” prominent tech analyst TP Huang said in a post on social media platform X on Tuesday, referring to computing cluster products Huawei unveiled domestically last year and overseas in March.
“Near frontier performance, trained on 50k Chinese domestic accelerators,” said Hanchi Sun, a computer science PhD student at Lehigh University focusing on LLM research, on X. “The first ever to achieve this!”
LongCat-2.0 boasted strong capabilities in coding and agentic tasks, according to Meituan.
The model outperformed Google’s older Gemini 3.1 Pro on several benchmarks, including Terminal-Bench 2.1 and SWE-Bench Pro, Meituan said, although it acknowledged that LongCat-2.0 still lagged behind global frontier models such as OpenAI’s GPT-5.5 and Anthropic’s Claude 4.8 Opus.
Meituan’s new model has not yet been evaluated on leading benchmarks such as Artificial Analysis and Arena, or cutting-edge tests such as Agents’ Last Exam and CyberGym.
Despite the milestone, Meituan said it still faced steep hurdles in bypassing Western graphics processing units (GPUs).
“Compared to the mature Nvidia GPU ecosystem, the supporting software community is still less developed,” the company said in a technical report accompanying the launch.
Pre-training the model on a cluster of more than 50,000 chips presented “significant system-level challenges due to both model and cluster scale”, the company said, adding that memory was the “primary bottleneck”.
Meituan said its domestic accelerators possessed significantly less memory per device than Nvidia’s H800 chip, which is banned from export to China under US rules.
To overcome these constraints, the delivery giant said it had to put “significant effort into building a stable, secure, and scalable infrastructure”, deploying a range of optimisation.


Philippine tribe vows to stand ground against US-led Pax Silica: ‘listen to us’

As the US weighs restrictions on cloud computing, how will China’s AI sector adapt?

Chinese AI chipmaker Hygon plots expansion from data centres to robotics

For China’s hi-tech IPOs, it’s Hong Kong vs the mainland – but how much does it matter?

Artificial intelligence (AI)

China claims biggest AI model trained on local chips, as Meituan releases LongCat-2.0

Meituan said the model was built entirely on “large-scale clusters of tens of thousands of AI ASIC superpods". Photo: Shutterstock Images

Faster AI, lower costs: DSpark eases inference bottlenecks and chip strain, says DeepSeek

DeepSeek says DSpark delivers faster results and lower costs, marking a breakthough in inference. Photo: Reuters

The new technique could reduce the computing resources needed to serve AI systems, according to a programmer. Shutterstock

Can Chinese silicon replace Nvidia? Here are 5 AI models trained on local chips

Chinese AI models have become increasingly competitive with their US peers, but the country’s AI hardware still lags significantly. Photo: Shutterstock Images

Huawei’s Ascend 910 processor on display at PT Expo China 2023. Photo: Shutterstock Images

Visitors at the Moore Threads booth at the World Artificial Intelligence Conference in Shanghai, July 4, 2024. Photo: CFOTO/Future Publishing via Getty Images

As AI pushes data centres to breaking point, some Chinese chipmakers bet on SiC

A view of Exchange Square in Central. Photo: Jelly Tse

Basic is expected to debut on public markets with an IPO in Hong Kong. Photo: Jelly Tse

China’s Zhipu AI sparks new ‘DeepSeek moment’ with cost-effective coding model

Zhipu AI, known internationally as Z.ai, made the GLM-5.2 model available on June 13. Photo: Shutterstock

The DeepSeek logo and words reading ‘Artificial Intelligence AI’ are seen in this illustration taken on January 29, 2025. Photo: Reuters

The Qwen and Alibaba logos are seen in this illustration taken January 29, 2025. Photo: Reuters

Tech

Tech Trends

LongCat-2.0 boasts 1.6 trillion parameters and a million-token context window, on par with DeepSeek’s latest flagship model


LongCat-2.0 boasts 1.6 trillion parameters and a million-token context window, on par with DeepSeek’s latest flagship model.


China claims biggest AI model trained on local chips, as Meituan releases LongCat-2.0

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}LongCat-2.0 boasts 1.6 trillion parameters and a million-token context window, on par with DeepSeek’s latest flagship model

LongCat-2.0 boasts 1.6 trillion parameters and a million-token context window, on par with DeepSeek’s latest flagship model