- LongCat-2.0 contains 1.6 billion parameters and a context of one million tokens
- Meituan trained the model using more than 50,000 domestic AI accelerators.
- The model completed pre-training completely without any Nvidia hardware involved.
Meituan has released LongCat-2.0, an open source large language model containing 1.6 trillion parameters and supporting a context window of 1 million tokens.
This scale puts the model roughly on par with DeepSeek’s flagship V4-pro, which launched in April this year.
Meituan says LongCat-2.0 has completed full-process training on a computing cluster containing more than 50,000 domestic AI accelerators, making it the first trillion-parameter model to reach that scale.
Home hardware reaches a new milestone in training
The announcement comes as China continues to expand its domestic computing capabilities amid export restrictions that limit access to advanced graphics processors from the United States.
Unlike DeepSeek V4-pro, which relied on Chinese chips only during inference, LongCat-2.0 also completed the much more demanding pre-training stage using domestic hardware.
This means that the company has completely avoided using foreign AI hardware like that from Nvidia.
The company said the system was built entirely on large AI ASIC superpods while using Huawei’s Collective Communication Library to improve the stability of communication between the processors.
China’s homegrown AI chips have been widely adopted for model inference amid Beijing’s push toward technological self-sufficiency, although pre-training remains significantly more difficult.
Meituan claims that LongCat-2.0 showed strong performance in coding and agent-based tasks, while outperforming Google’s Gemini 3.1 Pro in several benchmarks, including Terminal-Bench 2.1 and SWE-Bench Pro.
However, it acknowledged that its latest model still lags behind OpenAI’s GPT-5.5 and Anthropic’s Claude 4.8 Opus in broader frontier capability assessments.
“This put to rest any concerns about the Atlas-950 SuperPoDs. [being] unable to form large LLMs to [Zhipu AI] and DeepSeek,” said technology analyst TP Huang.
Technical hurdles persist despite higher ambitions
Despite the successes recorded by Meituan, this is not achieved without the significant hurdle of replacing Nvidia hardware.
The company faced significant engineering difficulties during development despite completing training without relying on restricted foreign graphics processors.
Meituan said memory became the main bottleneck because each domestic accelerator offered substantially less capacity than Nvidia’s H800 chip, which remains unavailable for export to China under U.S. rules.
Therefore, engineers created additional optimization systems aimed at maintaining stable, secure, and scalable training across the cluster despite its considerable size and complexity.
Hanchi Sun, a PhD researcher in computer science, described the achievement by writing: “Near-frontier performance, trained on 50,000 Chinese domestic accelerators,” before adding: “The first to achieve this!”
LongCat-2.0 has not yet appeared in any major independent evaluations, including Artificial Analysis, Arena, Agents’ Last Exam, or CyberGym, leaving external verification of several reported capabilities pending.
However, the statement suggests that Chinese developers are trying to reduce dependence on Nvidia by expanding domestic hardware beyond inference into large-scale training.
Broader comparative results between AI tools will ultimately determine how competitive this approach will be in the future.
via SCMP
Follow TechRadar on Google News and add us as a preferred source to receive news, reviews and opinions from our experts in your feeds.
