1 hour ago
29
Meituan, the Chinese company best known for delivering your dinner, just dropped one of the largest open-source coding models ever built. LongCat-2.0 packs 1.6 trillion parameters into a Mixture-of-Experts architecture designed specifically for agentic coding tasks, and it was trained entirely on domestically produced Chinese chips.
What LongCat-2.0 actually is
LongCat-2.0 is a Mixture-of-Experts model, which means it doesn’t 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. The model is enormous, but it only uses the parts it needs for any given task, 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.
On benchmarks, the model posted a score of 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench, evaluation suites designed to test how well AI models handle real-world software engineering challenges.
The model weights and associated resources are publicly available on Hugging Face under the meituan-longcat organization.
The hardware story underneath
Meituan trained and ran this model on a 50,000-card domestic compute cluster. That means no Nvidia A100s or H100s, no AMD MI300X chips. The entire training pipeline relied on Chinese-manufactured hardware.
US export controls have been systematically restricting China’s access to cutting-edge AI chips since late 2022. Meituan trained a 1.6 trillion parameter model without any of the restricted hardware. Meituan claims the 50,000-card cluster represents a first of its scale for domestic Chinese hardware.
The LongCat lineage
LongCat-Flash, a 560 billion parameter model, came out in September 2025. That was 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.
What this means for the AI landscape
For developers and researchers, the availability on Hugging Face lowers the barrier to experimentation. Fine-tuning a model of this caliber for specific use cases, whether that’s enterprise codebases, security auditing, or automated debugging, becomes a real option for teams that couldn’t afford to train something comparable from scratch.
Every model trained successfully on domestic Chinese hardware weakens the leverage of US export controls. If Chinese companies can produce competitive frontier models without American chips, the primary tool the US has used to maintain its AI advantage starts looking like a strategy with a shelf life.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·