- LongCat-2.0 contains 1.6 trillion parameters and a million-token context
- Meituan trained the model using over 50,000 domestic AI accelerators
- The model completed pretraining without any Nvidia hardware involved
Meituan has released LongCat-2.0, a large open source language model that contains 1.6 trillion parameters and supports a context window of 1 million tokens.
This scale places the model roughly on par with DeepSeek’s flagship V4-pro, which was launched back in April this year.
Meituan says LongCat-2.0 completed full process training on a computer cluster containing more than 50,000 domestic AI accelerators, making it the first trillion-parameter model to achieve that scale.
Domestic hardware reaches a new educational milestone
The announcement comes as China continues to expand domestic computing capacity amid export restrictions that limit access to advanced U.S. graphics processors.
Unlike DeepSeek V4-pro, which only relied on Chinese chips during inference, LongCat-2.0 also completed the far more demanding pre-training phase using domestic hardware.
This means that the company has completely avoided the use of foreign AI hardware such as those from Nvidia.
The company said the system was built entirely on large AI ASIC superpods, while using Huawei’s Collective Communication Library to improve communication stability across processors.
China’s homegrown AI chips have been widely used for model inference amid Beijing’s push for technology independence, although prior training has remained considerably 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 across several benchmarks, including Terminal-Bench 2.1 and SWE-Bench Pro.
Nevertheless, it acknowledged that its latest model still trails OpenAI’s GPT-5.5 and Anthropic’s Claude 4.8 Opus on broader cross-border capability ratings.
“This mitigated any concerns about the Atlas-950 SuperPoDs [being] unable to train large LLMs for [Zhipu AI] and DeepSeek,” said technology analyst TP Huang.
Despite greater ambitions, technical hurdles remain
Despite the successes recorded by Meituan, this does not come without the significant hurdle of replacing Nvidia hardware.
The company faced major technical difficulties during development despite having completed training without relying on limited foreign graphics processors.
Meituan said memory became the primary bottleneck because each domestic accelerator offered significantly less capacity than Nvidia’s H800 chip, which remains unavailable for export to China under US regulations.
Engineers therefore built additional optimization systems intended to maintain 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 limit performance, trained on 50,000 Chinese domestic accelerators,” before adding, “The first ever to achieve this!”
LongCat-2.0 has not yet appeared in major independent evaluations, including Artificial Analysis, Arena, Agents’ Last Exam, or CyberGym, leaving external verification of several reported features outstanding.
However, the release suggests that Chinese developers are trying to reduce Nvidia dependence by expanding domestic hardware beyond the ends for large-scale training.
Broader benchmark results across AI tools will ultimately determine how competitive this approach will be going forward.
Via SCMP
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews and opinions in your feeds.
