Chinese technology company Meituan has open-sourced LongCat-2.0, a 1.6-trillion-parameter mixture-of-experts AI model trained entirely on domestic Chinese AI chips — becoming the first publicly confirmed model at this scale to complete both pre-training and inference without any Nvidia hardware. The release marks a landmark moment in China’s effort to build a sovereign AI stack capable of operating outside Western semiconductor supply chains.
What Happened
LongCat-2.0 is a mixture-of-experts model with 1.6 trillion total parameters, trained on a 50,000-card cluster of domestic Chinese AI application-specific integrated circuits. Meituan used Huawei’s Collective Communication Library to coordinate communication across that hardware cluster during training, signalling Huawei’s deep involvement in the underlying infrastructure. The model has been released under a permissive open-source licence and is available for developers worldwide to download, fine-tune, and deploy.
Performance benchmarks place LongCat-2.0 near the global frontier. On SWE-bench Pro — an agentic coding benchmark that evaluates a model’s ability to resolve real-world GitHub issues autonomously — LongCat-2.0 scored 59.5, surpassing GPT-5.5’s published score of 58.6. The model also supports a one-million-token context window, enabling it to process entire codebases or lengthy technical documents in a single pass. Meituan designed it primarily as a reasoning engine for agentic coding tools and autonomous enterprise workflow automation.
Why It Matters
The significance of LongCat-2.0 is less about any single benchmark score and more about what the training run proves at a systems level. Previous Chinese frontier models — including DeepSeek’s V4-pro, which drew significant global attention — relied on domestic chips for inference while still depending on foreign silicon for the computationally intensive pre-training phase. LongCat-2.0 is the first publicly disclosed model at this parameter scale to close that gap, completing training from scratch on entirely domestic hardware.
That distinction carries substantial strategic weight. US export controls have restricted sales of Nvidia’s A100 and H100 series chips to Chinese companies since late 2022 and have tightened further in subsequent rounds. The AI industry has debated for years whether these restrictions would meaningfully delay China’s AI capability development. LongCat-2.0 offers the clearest counter-evidence yet. Custom silicon development is accelerating across the global AI industry, and China’s domestic ASIC ecosystem has now demonstrated it can support pre-training runs at the trillion-parameter scale.
LongCat-2.0’s open-source release also feeds directly into the competition for global developer mindshare. Xiaomi’s MiMo V2.5 recently topped OpenRouter traffic charts, the first time a Chinese model had led that platform’s leaderboard. LongCat-2.0’s arrival — stronger on coding benchmarks than MiMo V2.5 and now openly available for fine-tuning — extends the push by Chinese AI labs into the agentic reasoning and software engineering categories where OpenAI, Anthropic, and Google are competing most fiercely.
Background and Context
Meituan is best known internationally as a food delivery and local commerce platform, but the company has invested aggressively in AI research and autonomous systems over the past several years, with drone delivery, robotic kitchen logistics, and enterprise productivity tools all part of its portfolio. The development of a frontier-scale AI model reflects a broader pattern among Chinese internet conglomerates — Alibaba, Baidu, Tencent, and ByteDance have all launched large language model programmes — but Meituan’s approach is notable for its focus on fully domestic hardware from the outset.
China’s domestic chip ecosystem has advanced rapidly in response to export restrictions. Huawei’s Ascend AI accelerators, Cambricon’s training chips, and a growing number of specialised ASIC makers have collectively assembled a hardware supply chain that, while still trailing Nvidia’s peak compute efficiency per chip, has now proven capable of sustaining frontier-scale training runs across tens of thousands of cards.
What Comes Next
The open-source release makes LongCat-2.0 immediately available to research labs, enterprise developers, and startups globally. Meituan is expected to integrate the model into its own agentic product offerings, particularly enterprise workflow automation tools the company has been developing for the Chinese corporate market. International partners and cloud providers may also offer fine-tuned versions optimised for specific verticals.
Beyond the immediate product implications, LongCat-2.0 will inform ongoing policy debates in Washington over whether semiconductor export controls are achieving their intended strategic objectives. With a major Chinese tech company now demonstrating the ability to train frontier-class models entirely on domestic hardware, some analysts argue that restrictions are failing to delay China’s AI capability development while inadvertently accelerating the build-out of a domestic chip industry. Policymakers weighing further rounds of export controls will be studying this release carefully.
