China's artificial intelligence ambitions took a significant step forward on Tuesday when Meituan, the Beijing-based on-demand services giant better known for food delivery, unveiled LongCat-2.0, a model the company says represents a definitive proof of concept for large-scale AI development on Chinese-made hardware. The company open-sourced LongCat-2.0, a new large language model boasting 1.6 trillion parameters and a context window of 1 million tokens, placing it on par with DeepSeek's latest flagship model, V4-pro, which launched in April. The release carries geopolitical weight well beyond its technical specifications, as it directly challenges the assumption that frontier-scale AI training requires access to advanced Nvidia processors now blocked from export to China by the United States government.

The distinction that makes LongCat-2.0 significant is not merely its scale but what it was trained on. While DeepSeek-V4-pro relied on home-grown chips only for inference, the process where a pre-trained model runs to answer user queries, LongCat-2.0 used domestic hardware for both inference and pre-training, according to Meituan. Pre-training is a far more computationally intensive process, during which an AI model digests massive datasets to learn basic patterns. The hardware stack reportedly includes technology associated with Huawei Technologies, including the Huawei Collective Communication Library and Atlas-950 SuperPods, though Meituan did not name its hardware vendor explicitly. The company said the model was trained end-to-end on a 50,000-chip domestic compute cluster, a scale that independent researchers have called unprecedented for Chinese-built accelerators.

Beneath the headline numbers, LongCat-2.0 carries an architectural design built specifically for agentic coding, AI that can autonomously plan, write, and execute software across complex multi-step tasks. It is a Mixture-of-Experts model, meaning it does not fire all 1.6 trillion parameters every time it processes a piece of text. Instead, it dynamically activates roughly 33 billion to 56 billion parameters per token, making it far more efficient than a dense model of comparable size would be. The context window stretches to 1 million tokens, roughly the equivalent of feeding the model an entire large codebase and asking it to reason about the whole thing at once. In benchmark testing, LongCat-2.0 registered a score of 59.5 on SWE-bench Pro, surpassing GPT-5.5's benchmark of 58.6, and achieved 70.8 on Terminal-Bench 2.1 and 77.3 on SWE-bench Multilingual. The model still trails the very top global systems, including OpenAI's GPT-5.5 and Anthropic's Opus 4.8 across the most demanding agentic and reasoning tasks, though the gap has narrowed materially.

One of the more striking aspects of the launch is that LongCat-2.0 was not, in fact, an unknown quantity at the time of its unveiling. Meituan officially unmasked the model as the computational engine behind "Owl Alpha," an anonymous stealth model that had spent the last two months commanding global developer charts on OpenRouter. During that period, the model processed around 10.1 trillion monthly tokens, averaging 559 billion tokens daily, recording 242% month-on-month growth and ranking among OpenRouter's global top three models. The fact that an unbranded Chinese model built on domestic silicon was already leading global developer adoption metrics before its identity was revealed underscores the pace at which the competitive landscape is shifting. Meituan has now released the model on GitHub, Hugging Face, and its own platform under a commercially permissive MIT License, enabling developers and enterprises to modify, redistribute, and integrate it into proprietary software.

The release also illuminates the rapid maturation of Meituan as an AI player. LongCat-Flash, a 560-billion-parameter model, was released in September 2025, followed by LongCat-Next, a multimodal variant, in March 2026. LongCat-2.0 landed on June 30, 2026, nearly tripling the parameter count of its predecessor in less than a year. The company has publicly committed to investing heavily in AI and domestic chip capabilities as part of a strategic pivot away from its origins as a logistics and delivery platform. The progression from LongCat-Flash to LongCat-2.0 in under twelve months suggests that Meituan's AI research division has moved rapidly up the capability curve while simultaneously solving the engineering challenges of training at scale on hardware that lacks the mature software ecosystem of Nvidia's platforms. Meituan acknowledged that its software ecosystem still lags Nvidia's mature GPU community and that memory limits were the primary bottleneck during pre-training, as domestic accelerators carry less memory per device than Nvidia's banned H800 chip.

The broader implications of LongCat-2.0 extend well beyond Meituan and reframe the debate around the effectiveness of US export controls as a tool for slowing China's AI development. The model's training on Chinese-made chips demonstrates an end-to-end domestic AI stack, from hardware through training to deployment, directly responding to US export controls that have restricted Chinese access to Nvidia's most advanced GPUs. Industry observers noted the significance immediately. Lehigh University researcher Hanchi Sun described it as the first model ever trained to near-frontier performance on 50,000 Chinese domestic accelerators, while venture partner Alvin Foo observed that if China can scale frontier training on local silicon at this level, the compute arms race is far more open than recent forecasts had projected. For the broader AI market, LongCat-2.0 validates that Chinese labs can produce competitive large models on domestic silicon, potentially accelerating sovereign AI initiatives in other regions and intensifying the debate about whether export controls are effective in slowing China's frontier-AI progress.

Key Takeaways

Historic hardware milestone: LongCat-2.0 is the first trillion-parameter AI model to complete both full pre-training and inference on a domestic chip cluster, a 50,000-card array of Chinese-made AI ASICs, without any reliance on restricted Nvidia GPUs.
Scale and architecture: With 1.6 trillion total parameters and a 1-million-token context window using a Mixture-of-Experts design, LongCat-2.0 is on par with DeepSeek V4-pro and outperforms GPT-5.5 on key software engineering benchmarks.
Stealth market validation: The model was already operating anonymously as "Owl Alpha" on OpenRouter for two months before its unveiling, accumulating 10.1 trillion monthly tokens and ranking among the platform's global top three models.
Open-source strategy: Released under a permissive MIT License on GitHub and Hugging Face, LongCat-2.0 follows the open-source playbook that brought DeepSeek global prominence, maximizing developer adoption and commercial integration.
Geopolitical signal: The release challenges the core premise of US chip export controls, that restricting access to Nvidia hardware will meaningfully slow China's ability to train frontier AI models, and may accelerate sovereign AI strategies worldwide.

Loading Article...

Loading Article...