China's AI Chip Breakthrough
Meituan Launches LongCat-2.0, China's Largest AI Model Trained on Domestic Chips
Anonymous
Key Points
1
LongCat-2.0's benchmark results show SWE-Bench Pro 59.5, Terminal-Bench 2.1 70.8, and SWE-Bench Multilingual 77.3, placing it ahead of GPT-5.5 on SWE-Bench Pro.
2
Before its public release, LongCat-2.0 ran anonymously on OpenRouter for about two months under the codename 'Owl Alpha' and was released via GitHub, Hugging Face, and Meituan's platform.
3
Meituan claims LongCat-2.0's performance is comparable to Google's Gemini 3.1 Pro, signaling parity with a leading international model.
4
Training reportedly used a domestic cluster of more than 50,000 ASIC chips and was conducted without Nvidia GPUs.
5
LongCat-2.0's API pricing is $0.75 per 1 million input tokens and $2.95 per 1 million output tokens, with a promotional period offering $0.30 input and $1.20 output tokens.
Meituan, China's food delivery giant, has released LongCat-2.0 — a 1.6-trillion-parameter AI model trained entirely on domestic chips, with no Nvidia hardware involved. Reuters reports the model runs on a cluster of more than 50,000 Chinese-made ASICs and can process inputs of up to one million tokens at a time.
The release lands as a direct challenge to U.S. export controls meant to slow China's AI ambitions. LongCat-2.0 scores 59.5 on SWE-Bench Pro, edging out GPT-5.5's 58.6, according to VentureBeat. Meituan open-sourced the model under a permissive MIT license, making it free for commercial use.
Secret 'Owl Alpha' Was LongCat-2.0 All Along
Before today's launch, LongCat-2.0 ran anonymously on the AI platform OpenRouter for about two months under the codename "Owl Alpha," according to VentureBeat. Developers had no idea who built it. By May 2026, Owl Alpha had climbed to a top-3 ranking on OpenRouter, processing 10.1 trillion tokens per month.
Meituan unmasked Owl Alpha as LongCat-2.0 on June 30, releasing model weights via GitHub and Hugging Face. Cybernews notes the model is available for anyone to download and use. Developers on Hacker News praised the MIT license, which lets companies build proprietary products on top of it without any forced disclosure.
Trained on 50,000 Chinese Chips, Not One Nvidia GPU
The core claim is significant: Meituan says LongCat-2.0 completed both pre-training and inference on domestic ASICs. Reuters and South China Morning Post both note the company used the Huawei Collective Communication Library (HCCL) to keep the massive chip cluster stable during training. HCCL is software that helps thousands of chips work together without errors.
Previous Chinese models like DeepSeek had used domestic chips only for inference — the easier, cheaper phase. Pre-training, which is far more demanding, had still relied on foreign hardware. Meituan's end-to-end domestic run is what sets LongCat-2.0 apart, according to The Next Web.
Benchmark Numbers Put It Ahead of GPT-5.5 on Coding
LongCat-2.0 scores 59.5 on SWE-Bench Pro, a coding benchmark, beating GPT-5.5 (58.6) and Gemini 3.1 Pro (54.2), per VentureBeat. On Terminal-Bench 2.1, it scores 70.8 — nearly matching Gemini 3.1 Pro's 70.7. On SWE-Bench Multilingual, it hits 77.3, comparable to Claude Opus 4.6.
The model uses 1.6 trillion total parameters but activates only around 48 billion per token. That design — called Mixture-of-Experts — keeps costs low while maintaining high performance. Meituan trained it on more than 35 trillion tokens of data, according to Yahoo Finance.
Aggressive Pricing Signals a New Chinese API Price War
Meituan set its standard API price at $0.75 per million input tokens and $2.95 per million output tokens. During a promotional period, those prices drop to $0.30 and $1.20. Context caching is free. That is roughly 60% cheaper than standard rates, a cut that South China Morning Post says could spark a new price war among rivals like Alibaba and DeepSeek.
Meituan CEO Wang Xing framed the launch in sweeping terms. "In the AI revolution, the only reasonable strategy is to go on the offensive," he said. The company plans to plug LongCat-2.0 into its "Xiao Tuan" assistant, turning its food delivery app into a platform that can autonomously book travel and manage merchant logistics, per Cybernews.
Sentiment
Mostly PositiveNEGATIVENEUTRALPOSITIVE
Publishers
16
Articles
51
Reach
67
Topics
A
AnonymousAuthor
