Meituan's LongCat-2.0: First trillion-parameter AI model trained entirely on Chinese chips - open source under MIT license
What it really says
Chinese technology company Meituan officially unveiled its AI model LongCat-2.0 on June 30, 2026. With 1.6 trillion parameters, it is the first model of this scale to be trained entirely on domestically produced Chinese semiconductors, using a cluster of 50,000 locally manufactured processors without any Nvidia GPUs or other Western chips. The model uses a Mixture-of-Experts architecture: of the 1.6 trillion total parameters, only between 33 and 56 billion are activated per processing step (averaging approximately 48 billion), which dramatically reduces computational costs. The native context window spans one million tokens. LongCat-2.0 is designed as an agentic coding model and scores 59.5 on SWE-bench Pro, which measures the ability to resolve real GitHub issues across complete repositories. It achieves 70.8 on Terminal-Bench 2.1, 77.3 on SWE-bench Multilingual, and 73.2 on the corporate workflow simulator FORTE. The model had previously been leading usage statistics on the OpenRouter platform under an internal codename before its identity was revealed. Meituan is releasing the model under the permissive MIT license and offering it at approximately $0.75 per million input tokens and $2.95 per million output tokens, significantly cheaper than comparable Western models. However, the full model weights have not yet been published; only the announcement is available on GitHub and Hugging Face.
Our assessment
This model merits a yellow rating because it marks a geopolitically significant shift whose implications require nuanced consideration. The concerning side: LongCat-2.0 demonstrates that US chip export controls are failing to achieve their strategic objective. Despite massive restrictions on access to Nvidia GPUs and other high-performance chips, a Chinese company has trained a near-frontier model on purely domestic hardware. This means the notion that global AI development can be slowed through chip controls is outdated. The AI race is accelerating, and no single country or alliance can control it alone. Furthermore, the model is released under the MIT license, meaning it can be used worldwide without restrictions, including for applications that Western providers would exclude through their terms of service. The reassuring side: open models enable independent safety research and reduce dependency on individual providers. The Mixture-of-Experts architecture shows that powerful AI does not necessarily have to consume ever more energy, as only a fraction of parameters are activated per query. And competition drives down prices: when a model of this class is available for under $3 per million tokens, European companies and developers ultimately benefit as well.
Relevance for Germany
This development carries strategic significance for Germany. First, the German government and the EU base their AI strategy on the assumption that Western technological leadership can be secured through chip controls. LongCat-2.0 disproves this assumption and forces a rethink: it is not enough to control hardware access when algorithms can compensate for hardware disadvantages. Second, German companies using AI models gain another powerful alternative to US providers with LongCat-2.0. After the Fable 5 export ban in June, which showed how quickly access to US AI services can be restricted, the issue of technological sovereignty becomes even more pressing in Berlin. Third, the model is optimized for agentic coding, precisely the area that industry experts predict will trigger the next major labor market transformation. Fourth, the aggressive pricing of under $3 per million tokens puts additional competitive pressure on the European AI industry.
Fact check
The release of LongCat-2.0 on June 30, 2026 is consistently reported by VentureBeat, South China Morning Post, KuCoin News, and CryptoBriefing. The technical specifications - 1.6 trillion parameters, Mixture-of-Experts with 33-56 billion active parameters, one million token context window - are consistently documented across all sources. The benchmark results (SWE-bench Pro 59.5, Terminal-Bench 2.1: 70.8, SWE-bench Multilingual: 77.3, FORTE: 73.2) are confirmed by multiple sources. Training on a cluster of 50,000 domestically manufactured chips without Nvidia hardware is consistently reported by VentureBeat and SCMP. The MIT license and pricing (approximately $0.75 and $2.95 per million tokens) are confirmed in multiple reports. That the model weights had not yet been published at the time of the announcement is noted on the Hugging Face and GitHub pages.
Source
- • https://venturebeat.com/technology/meituan-open-sources-longcat-2-0-the-1-6t-near-frontier-agentic-coding-model-thats-been-leading-openrouter-trained-entirely-on-chinese-chips
- • https://www.scmp.com/tech/tech-trends/article/3358854/china-debuts-biggest-ai-model-trained-local-chips-meituan-releases-longcat-20
- • https://www.kucoin.com/news/flash/meituan-launches-longcat-2-0-with-1-6-trillion-parameters-and-competitive-pricing
- • https://cryptobriefing.com/meituan-longcat-2-undercuts-gpt-claude-pricing/