Meituan open-sources 1.6 trillion-parameter AI model trained on 50,000 domestic chips

Meituan open-sourced LongCat-2.0, a 1.6 trillion-parameter large language model trained on more than 50,000 domestic AI accelerators, marking the first time a model of this scale completed pre-training without Nvidia hardware.

"This put to rest any concerns of Atlas-950 SuperPoDs being unable to train large LLMs," said TP Huang, a tech analyst covering Chinese AI infrastructure.

The model supports a 1 million-token context window, placing it roughly on par with DeepSeek's V4-pro, which launched in April. Meituan said it used Huawei's Collective Communication Library to maintain communication stability across the cluster. Memory was the primary bottleneck, as each domestic accelerator offered substantially less capacity than Nvidia's H800 chip, which remains unavailable for export to China under US export restrictions. Engineers built additional optimization systems to maintain stable, secure, and scalable training across the cluster.

Citi maintained a Buy rating on Meituan with a HKD113 target price, saying the open-source strategy strengthens the company's competitive moat in local services. The broker cited Meituan's proprietary offline transaction data, rider network data, and dispatch algorithm experience as differentiated advantages over competitors such as ByteDance. The stock rose 5% following the announcement.

Domestic hardware reaches a new training milestone

The achievement comes as China expands domestic computing capabilities amid US export restrictions limiting access to advanced graphics processors. Unlike DeepSeek V4-pro, which relied on Chinese chips only during inference, LongCat-2.0 completed the more demanding pre-training stage using domestic hardware. The company said the system was built entirely on large AI ASIC superpods.

Meituan claims LongCat-2.0 showed strong performance in coding and agent-based tasks, exceeding Google's Gemini 3.1 Pro on Terminal-Bench 2.1 and SWE-Bench Pro. The company acknowledged it still trails OpenAI's GPT-5.5 and Anthropic's Claude 4.8 Opus on broader frontier capability assessments. The model has not yet appeared on independent evaluations including Artificial Analysis, Arena, or CyberGym, leaving several reported capabilities unverified by third parties.

Hanchi Sun, a computer science PhD researcher, described the achievement as "near frontier performance, trained on 50k Chinese domestic accelerators — the first ever to achieve this."

Investment implications for AI infrastructure

The development challenges Nvidia's dominance in China's AI training market. Nvidia's H800 and H100 chips are restricted for export to China under US rules, creating an opening for domestic alternatives from Huawei and other Chinese chipmakers. Meituan's successful pre-training on 50,000 domestic accelerators suggests Chinese developers are reducing dependence on Nvidia hardware beyond inference into large-scale training.

For Meituan, the open-source release serves a dual purpose: attracting external developers to its platform while deepening merchant relationships through AI-driven marketing tools built on proprietary transaction data. Citi's analysis suggests this creates a competitive moat that differentiates Meituan from other large language model providers in China's local services market.

Broader benchmark results across AI evaluation tools will determine how competitive the domestic hardware approach becomes. Meituan shares trade at approximately 22x forward earnings, with the HKD113 target price implying about 30% upside from current levels.

This article is for informational purposes only and does not constitute investment advice.