Meituan unveils LongCat-2.0 AI model trained entirely on domestic Chinese chips

China's Meituan says new AI model trained on domestic chips

BEIJING, June 30 (Reuters) - China's food delivery giant Meituan (3690.HK), opens new tab said on Tuesday it had released and would open-source its next-generation LongCat large language model, claiming it is the world's first trillion-parameter AI system trained and run entirely on a 50,000-chip cluster powered by Chinese-made processors. Meituan, often compared to DoorDash, is a late entrant to China's crowded and well-funded AI sector, where rivals include DeepSeek and ⁠ByteDance's Doubao. The LongCat team, founded in 2023, only launched its first model late last year. Although the company did not disclose how the new model, LongCat-2.0, will be integrated into its existing businesses, Meituan has used earlier versions to power in-app AI assistants that recommend restaurants and hotels and complete tasks such as ordering food and booking rooms, part of an "agentic commerce" trend rival Alibaba has accelerated this year. Amid weak consumer sentiment and diminishing margins, Meituan may also be seeking to diversify revenue streams. In a statement on LongCat's official WeChat account, the company highlighted the model's ability to build a gaming website and write a novel. SELF-SUFFICIENCY FOCUS LongCat-2.0's reliance on Chinese AI chips underscores the ⁠growing importance of self-sufficiency in China's domestic AI market, as DeepSeek, Alibaba (9988.HK), opens new tab, ByteDance and other major players work to reduce dependence on U.S. chips to train their models, following export controls imposed by Washington since 2022. Chipmakers, including Huawei and Enflame, have moved quickly to fill the gap left by U.S. chipmakers, gaining market share through supply deals with AI developers. LongCat-2.0 was ⁠trained from scratch using 50,000 domestic chips and can process inputs of up to 1 million tokens, allowing it to handle ultra-long documents, according to the statement. The model is aimed at agentic coding, with its architecture designed to help it handle real-world ⁠coding tasks more efficiently and reliably. Ad Break Coming Up NEXT StayNext OffEnglish 180p288p360p480p540p576p720pHD1080pHDAuto (180p) About ConnatixV492702563 About ConnatixV492702563 1/1 Skip Ad Continue watchingafter the adVisit Advertiser websiteGO TO PAGE A preview version of the model had already become one of the three most-used models on OpenRouter, a globally popular AI marketplace, the company said. LongCat-2.0 matched or exceeded several leading proprietary models, ⁠including Google's Gemini, OpenAI's GPT-5.5 and Anthropic's Claude Opus, on some coding and agent benchmarks, Meituan claimed. "LongCat-2.0 has demonstrated that we now have the capability to train large-scale models on domestic computing clusters," the Chinese tech giant said, without naming the chipmaker. Reporting by Ethan Wang and Eduardo Baptista, Editing by Louise Heavens Our Standards: The Thomson Reuters Trust Principles., opens new tab * Suggested Topics: * Disrupted Eduardo Baptista Thomson Reuters Eduardo Baptista is a Senior Correspondent for Reuters based in Beijing, covering China's technology, space, and automotive industries. He has led enterprise and investigative reporting on China's military-linked companies, artificial intelligence and semiconductor supply chains, as well as macroeconomic and industrial policy. Baptista has reported from China for nearly a decade and holds a BA in History from the University of Cambridge.

Meituan open sources LongCat-2.0, the 1.6T, near-frontier agentic coding model that's been leading OpenRouter -- trained entirely on Chinese chips

A few hours ago, Chinese delivery app company Meituan officially unveiled LongCat-2.0 on GitHub, Hugging Face, and its native platform, unmasking the model as the computational engine behind "Owl Alpha," the anonymous stealth model that has spent the last two months commanding global developer charts on OpenRouter. Developed to fundamentally disrupt closed-source enterprise dominance in autonomous software engineering, the 1.6-trillion-parameter Mixture-of-Experts (MoE) system brings a native 1-million-token context window to the public domain under a highly permissive, enterprise grade, commercially viable MIT license. Commercial access to the architecture introduces a highly aggressive pricing tier, deploying a mechanism where all context-cache hits are processed completely free of charge, running alongside a time-limited "Token Pack" flash-sale paradigm. There's also a typical "pay-as-you-go" API for non-cache hits standard priced at $0.75/$2.95 per million tokens in/out. However, a limited-time promotional discount aggressively slashes these operational expenditures down to $0.30 per million tokens for uncached input and $1.20 per million tokens for output, both on the cheaper-end of top performing models globally. What makes the release a definitive inflection point for global tech infrastructure is its operational independence: the massive model was trained entirely on a cluster of over 50,000 domestic Chinese Application-Specific Integrated Circuits (ASICs), proving that near-frontier AI models can be scaled successfully without relying on the typical U.S. Nvidia GPUs that have, to date, powered much of the global generative AI frontier model training effort. This successful deployment of alternative silicon signals a profound structural shift. If Chinese conglomerates can consistently iterate trillion-parameter architectures using homegrown ASICs rather than general-purpose GPUs, it would seem to threaten Nvidia's dominance in this sector. Crucially, this technological pivot arrives precisely as Washington pressures top-tier American labs to restrict access to their latest models. Following a U.S. governmental request, OpenAI was forced to limit access to its new GPT-5.6 models, while Anthropic was previously also ordered by the U.S. to restrict access to its latest Claude Fable 5 / Mythos 5 models, which it took entirely offline in response. At the same time, a growing chorus of technologists, activists, and industry experts warn that these defensive regulatory maneuvers have inadvertently backfired. By locking down Western closed-source models and driving up API costs, the U.S. government has left a wide operational window for global developers seeking affordable, high-performance alternatives like those found in Chinese open source models such as Meituan LongCat-2.0. The raw operational metrics backed up the developer enthusiasm: during its unbranded residency on OpenRouter, Owl Alpha accounted for approximately 10.1 trillion monthly tokens -- averaging 559 billion tokens per day -- representing a 242% month-over-month explosion in volume that propelled it into the platform's global top three. By the time Meituan stepped forward to claim the architecture, the model had already secured the top ranking on the Hermes Agent workspace, second place on Claude Code deployments, and third place across international OpenClaw environments. Technology: Engineering the 1M-Token Sparse Context At the core of LongCat-2.0 lies an aggressive optimization of Mixture-of-Experts (MoE) sparsity, scaling total parameters to 1.6 trillion while limiting active computation to an average of 48 billion parameters per token. Depending on the structural complexity of a query, the model's dynamic activation ranges from 33 billion to 56 billion parameters. This design implements a "Zero-Compute Experts" framework, ensuring that routine execution elements pass through lighter subnetworks, entirely eliminating the idle computational overhead that typically penalizes ultra-dense models. To sustain a functional 1-million-token context window without incurring catastrophic hardware bottlenecks, Meituan introduced LongCat Sparse Attention (LSA). Designed as an evolutionary iteration of DeepSeek Sparse Attention, LSA resolves the quadratic scoring costs and memory fragmentation that typically plague fine-grained sparse mechanisms through three distinct, orthogonal vectors: * Streaming-aware Indexing (SI): This system restructures the token selection pipeline by blending hardware-aligned contiguous data reads with dynamic random selection. By converting fragmented memory access into highly predictable, sequential blocks, the system achieves coalesced High Bandwidth Memory (HBM) utilization and elevated effective bandwidth. * Cross-Layer Indexing (CLI): Leveraging the empirical reality that attention saliency remains highly stable across adjacent hidden layers, CLI amortizes calculation costs. A single indexing pass successfully guides multiple consecutive layers during inference, a capability reinforced by cross-layer distillation throughout the training phase. * Hierarchical Indexing (HI): This approach applies a coarse-to-fine, two-stage scoring layout. The indexer performs a rapid, approximate block-level recall to filter candidates, before running fine-grained token selection exclusively on the remaining population. Furthermore, Meituan integrated an N-gram Embedding module inherited from its lighter model lines. By expanding parameter allocation in sparse dimensions completely orthogonal to the MoE expert layout, the architecture appends 135 billion parameters to a 5-gram token combination framework. This expands the core embedding space by roughly 100-fold, allowing the model to capture dense local token relationships and accelerate large-batch inference operations by reducing memory Input/Output (I/O) bottlenecks. Product: Post-Training, MOPD Framework and Benchmark Performance While generalist large language models prioritize fluid, conversational interfaces, LongCat-2.0 focuses explicitly on multi-step engineering tasks, tool integration, and automated repository manipulation -- agentic tasks, in other words. In standardized assessments, LongCat-2.0 registers an empirical 59.5 on SWE-bench Pro, surpassing GPT-5.5's benchmark of 58.6. The model further establishes its agentic specialization by marking a 70.8 on Terminal-Bench 2.1, a 77.3 on SWE-bench Multilingual, and a 73.2 on the general corporate workflow simulator FORTE. This precise operational behavior is achieved through a structural post-training layer called Multi-Teacher Optimization via Mixture of Specialized Experts (MOPD). Rather than blending raw human feedback into a singular reward function, the MOPD architecture segregates post-training optimization into three independent, highly focused expert clusters. * The Agent Experts are fine-tuned strictly for structural execution, specializing in precise tool invocation, multi-turn API parameter parsing, and self-correcting loop mechanisms to avoid execution stagnation. * The Reasoning Experts are optimized in isolation to advance multi-hop logic, complex chain-of-thought engineering, mathematics, and high-level STEM problem-solving. * The Interaction Experts focus entirely on human alignment, instruction-following nuances, factual grounding to suppress hallucinations, and maintaining rigid safety guardrails without diminishing the model's overall utility. By segregating these vectors during post-training, LongCat-2.0 prevents functional degradation. A dynamic gate-routing mechanism then seamlessly fuses these specialized behaviors at runtime, allowing the final model to coordinate deep reasoning, stable tool execution, and safe user interaction simultaneously While LongCat-2.0 generally trails premium frontier systems like Claude Opus 4.8 across broad general-agent benchmarks such as FORTE and BrowseComp, it explicitly punches above its weight in software engineering. What makes this open-weight architecture special is its hyper-focus on autonomous development; it manages to narrowly exceed OpenAI's proprietary GPT-5.5 on the rigorous software engineering benchmark SWE-bench Pro (scoring 59.5 against 58.6), proving it is highly capable and fiercely competitive for complex coding tasks despite a leaner computational footprint. Commercial Framework: Pay-As-You-Go vs. Flash-Sale Token Packs Meituan's deployment strategy introduces a specialized commercial model that splits network access between conventional real-time API billing and structured "Token Packs". For traditional enterprise integration, standard top-up accounts are available, deducting operational capital in real time based directly on token input and generation metrics. However, to accommodate the unpredictable compute bursts characteristic of autonomous development agents, Meituan launched a structured Token Pack framework. Purchased as fixed, one-time volumetric allocations valid for a strict 30-day window, these packages stack directly on top of an organization's existing baseline API account. To manage network load across its ASIC clusters, Meituan releases these high-volume packages via limited flash sales four times daily, precisely at 10:00, 16:00, 21:00, and 23:00 Beijing Time on a first-come, first-served basis.The economic standout of this framework is the zero-charge processing of context cache hits. In massive agentic environments where a coding assistant must repeatedly read, reference, and modify the same multi-million-token code repository over an extended session, standard architectures penalize developers by charging full pricing for repeated input context. Under Meituan's infrastructure, only cache-miss inputs and final token generations consume the package quota. This architecture completely alters the operational cost economics of large-scale agent software development, enabling deep iterative context exploration without compounding costs. Licensing: Open-Source Structural Freedom By registering the LongCat-2.0 repository under the open-source MIT License, Meituan positions the architecture with maximum legal flexibility for enterprise integration. In contrast to copyleft paradigms like the GNU General Public License (GPL) -- which legally obligates developers to open-source any derivative frameworks or internal software that links to the code -- the MIT license permits near-unrestricted freedom. For corporate engineering teams, this legal standard ensures that LongCat-2.0 can be deeply modified, compiled, and hard-coded directly into closed-source commercial applications, proprietary dev tools, and internal automation backends. Corporations can fork the repository, optimize the internal LSA mechanisms for private databases, and sell the resulting software stack to end users without any obligation to disclose their proprietary intellectual property or structural enhancements. Meituan's Evolution: From Delivery Super App to AI Powerhouse Founded in March 2010 by serial entrepreneur Wang Xing, Meituan initially launched as a Groupon-style daily deals website before rapidly evolving into one of China's dominant "super apps". Following a massive 2015 merger with Dianping, the Beijing-based tech giant solidified a dominant market share over the country's urban delivery corridors, bridging local consumer reviews, instant retail, hotel bookings, and food delivery. Operating as a publicly traded powerhouse on the Hong Kong Stock Exchange, Meituan claims over 770 million annual transacting users and supports a network of more than 14.5 million merchants. However, faced with intense domestic market competition, severe margin compression, and a sliding profit margin, the company aggressively pivoted its strategy beyond logistics. Meituan publicly committed to investing "billions" into artificial intelligence and domestic chip capabilities to revitalize its technology-driven offerings. This strategic shift into the global AI race began materializing in late 2025 with the release of LongCat-Flash, a 560-billion-parameter Mixture-of-Experts foundation model, followed quickly by the advanced reasoning model LongCat-Flash-Thinking. By open-sourcing these frontier-class models under enterprise-friendly licenses, Meituan signaled its ambition to become a foundational player in global AI infrastructure rather than remaining strictly a regional e-commerce and delivery giant. Enterprise Implications: Autonomous Operational Workflows For modern enterprises, the release of LongCat-2.0 unlocks clear operational strategies across software engineering, system operations, and long-form data interpretation. The combination of an open-weight, MIT-licensed model with an expansive 1-million-token context window means organizations can bypass the data privacy concerns and recurring overhead associated with hosting proprietary third-party APIs.In large-scale enterprise development environments, teams can leverage the model's specialized Agent Experts to orchestrate autonomous codebase migrations. Instead of dedicating hundreds of developer hours to manually rewriting legacy application frameworks, engineers can pass an entire enterprise repository along with modern SDK documentation directly into the 1-million-token context window. LongCat-2.0 can map the dependencies, execute the repository-level structural updates, compile the new codebase, and catch compilation and execution bugs autonomously within local sandbox environments before generating a final pull request. The model's architectural separation via the MOPD gate-routing mechanism yields significant advantages for strict enterprise compliance. By routing specific operational queries through isolated expert clusters, a financial institution or healthcare firm can deploy deep logic and mathematical reasoning passes without risking factual hallucination or violating strict safety bounds. The Interaction Experts function as an implicit guardrail layer, suppressing errors and enforcing instruction-following protocols without degrading the raw processing power of the internal Reasoning Experts. Combined with the zero-cost caching model, enterprises can maintain hyper-focused autonomous software networks that can repeatedly inspect corporate data pools, continuously maintaining and optimizing internal infrastructure at a fraction of standard operational costs.

China's Meituan says new AI model trained on domestic chips

Chinese tech firm Meituan launched a new artificial intelligence model on Tuesday that it said was the first of its size to be trained using domestically developed computer chips. The country is locked in a fierce race for AI dominance with the United States, which restricts exports to China of the most cutting-edge silicon chips designed by industry leader Nvidia. In response to Washington's restrictions, imposed on national security grounds, China has been accelerating efforts to develop its own advanced chips and break away from reliance on U.S. hardware. Meituan on Tuesday unveiled LongCat-2.0, a new large language model whose performance it says is comparable to Google's Gemini 3.1 Pro, released in February. LLMs are the kind of technology that underpins chatbots and many other AI tools. LongCat-2.0 is "the industry's first trillion-parameter model to complete end-to-end training and inference on a 50,000-chip domestic compute cluster," Meituan said in a statement. It did not disclose which Chinese chipmaker's products were used in the training process. The announcement marks a milestone for China's fast-developing AI industry because training competitive models with vast amounts of digital data requires chips with hefty computing power. Flagship models from other Chinese AI labs, including DeepSeek and Zhipu, can run inference on chips produced by domestic companies like Huawei -- a less intensive process than training. It is believed that most domestic models were trained using Nvidia chips. Some media reports have suggested that an unknown number of advanced Nvidia AI chips were smuggled into China. Meituan's AI research team began exploring the use of domestic chips in 2023 and "has proven that we are now capable of carrying out large-scale model training on domestic compute clusters," the company said.

LongCat-2.0: The Stealth AI Model That Was Quietly Topping OpenRouter All Along

Chinese tech company Meituan officially unveiled LongCat-2.0 on June 30, confirming the open-license, 1.6-trillion-parameter mixture-of-experts AI model is the same system that spent two months running anonymously on OpenRouter under the alias Owl Alpha. Parameters are the total number of dials a model can handle during training. The model activates roughly 48 billion of its parameters per token (the smallest unit of data an AI model processes), with that figure swinging between 33 billion and 56 billion depending on how demanding the query is. The stealth period paid off. By the time Meituan stepped forward, the model had already taken first place on the Hermes Agent workspace, second on Claude Code, and third across OpenClaw deployments, all ranked by monthly call volume. This is the first trillion-parameter model trained and deployed end-to-end on domestic Chinese ASICs, not just served on them after training elsewhere. DeepSeek's V4-Pro, by comparison, used Huawei chips only for inference while pretraining ran on Nvidia hardware. Meituan says the pretraining run, spanning more than 35 trillion tokens across a cluster of over 50,000 domestically produced accelerators, finished with "no rollbacks or irrecoverable loss spikes." That stability claim matters given how often large training runs on unproven hardware stacks fail midway through and how China seems to be reducing its dependence on U.S. hardware to train its models. Price is where LongCat-2.0 makes its real case. Standard API access runs $0.75 per million input tokens and $2.95 per million output, cut to $0.30/$1.20 during the current launch promo, with cached context reads free of charge. That undercuts GPT-5.5's $5/$30 per million tokens, Claude Sonnet 5's introductory $2/$10 rate, and lands close to DeepSeek V4-Pro's permanent $0.435/$0.87 and Xiaomi's MiMo-V2.5 Pro, which matched that same rate after its own May price cuts. Meituan also provides a token plan, which makes things even cheaper for coders and heavy users, offering packs of 1 billion tokens at around $60. We ran LongCat-2.0 through a quick game-building test ourselves. It got the job done, and the output held up reasonably well after a few rounds of iteration. The result landed visibly behind Claude Fable and Opus 4.8, making it easier to rank near Sonnet 4.6, but the quality-per-dollar math is hard to argue with at these prices. It made the waves of enemies come from different angles with the camera auto centering on the nearest enemy. However, the model's logic didn't take into consideration what happens when the number of enemies increases with difficulty. At higher speeds, the target-switching logic became erratic; the focus would jump to a closer enemy in the middle of a typing prompt, making the game frustratingly unplayable. This is normal in vibe coding sessions, where models don't foresee many logical consequences of a decision, and instead focus on delivering a result based on what the user prompts, literally. This is also why a cheap model is always a good option, because it gives the user more chance to iteratively improve every result until the final product meets expectations. If anything, without further interaction, at first glance the overall quality lands somewhere in between DeepSeel v4 Flash and Deepseek v4 Pro in our quick coding tests. You can check out the results in our itch.io site How Meituan built it LongCat-2.0 uses several techniques to make the model faster and more capable without dramatically increasing its size. Its attention system, based on DeepSeek's design, focuses only on the most relevant parts of very long conversations instead of processing everything equally, helping it respond more quickly. Also, a new N-gram embedding system (a way of helping understand groups of words or subwords together) gives the model a much richer understanding of words and phrases -- about 100 times more possible representations -- without adding many more AI components. It's basically teaching the AI to recognize common phrases instead of just individual words. Rather than seeing "New," "York," and "City" as three separate pieces, it can also treat "New York City" as a single meaningful concept. This gives the model a much richer understanding of language without making it dramatically larger. After training, Meituan also combines three specialized systems, one focused on using tools (Agent), one on solving problems (Reasoning), and one on conversations (Interaction). A routing mechanism then decides which combination of those specialists should handle each request, much like assigning the right team to the right job. On SWE-bench Pro, a benchmark that scores how often a model resolves real GitHub issues pulled from production codebases, LongCat-2.0 hit 59.5, ahead of GPT-5.5's 58.6 and Gemini 3.1 Pro's 54.2, though still behind Claude Opus 4.7 and 4.8. On FORTE, which grades agents on day-to-day office tasks across 15 professions under a 45-minute time limit, it scored 73.2, tied with Claude Opus 4.6 but trailing GPT-5.5's 77.8. Teams building coding agents on a budget, or anyone running high-volume repository-scale work where the free context-cache reads compound, get the clearest win. The model is reachable today through Meituan's OpenAI- and Anthropic-compatible API endpoints, or through agent harnesses like Hermes, Claude Code, and OpenClaw that already integrate it. Anyone who needs to self-host is out of luck for now. Both the GitHub and Hugging Face repositories still read "model weights coming soon," but Meituan has not set a date for when the files will ship.

China's Meituan open-sources massive LongCat-2.0 AI model, saying it was trained on domestic chips

China's Meituan open-sources massive LongCat-2.0 AI model, saying it was trained on domestic chips Beijing, China-based Meituan Inc. today debuted its next-gen LongCat-2.0 open-source large language model, stating that the company trained the 1.6-trillion-parameter model on domestic Chinese chips and compute clusters. The larger takeaway for this colossal model isn't just the open-source release, it's the domestic hardware throughline. Meituan may initially seem like an unlikely place for AI models to come from. Often described as China's answer to DoorDash, the company started as the country's dominant food delivery platform but evolved into a combination of services that include travel and leisure booking, discovery and rating of local businesses and ride-hailing. The company jumped into the AI model development scene as far back as 2023 when it acquired the startup Light Year Beyond for around $281 million, but did not announce its own internal plans to develop an AI model until 2025. At a high level, LongCat-2.0 follows a similar sparse Mixture of Experts model as Mistral AI's Mixtral and DeepSeek. It uses an internal router architecture that selects a curated set of "expert AIs" per token rather than lighting up the entire model at once. This provides core efficiency for model deployment and inference, allowing MoE models to scale on cheaper hardware without needing to deploy the entire model to compute every token. Weighing in at 1.6 trillion parameters, the model is no lightweight and it delivers with a 1-million-token context window. This means that users can input tremendous amounts of data at once. By way of comparison, it sits alongside MoE models such as DeepSeek-R1-0528 and OpenAI Group PBC's open-source GPT-OSS, which emphasize smaller activation footprints and industry-standard 128,000-token context windows, whereas LongCat-2.0 focuses on being very heavy and providing long context. The company has released benchmarks for the model pace it with ultra-powerful closed-source industry models such as Google LLC's Gemini, OpenAI's GPT-5.5 and Anthropic PBC's Claude Opus. The company said it designed LongCat-2.0 to work as a "brain," or the core, of AI agents and coding harnesses such as Claude Code, OpenClaw and Hermes. Meituan said the model delivers strong performance for code understanding, repository-level edits, automated task execution and agentic workflows. The objective is to provide developers with a stable and efficient tool that uses a model to orchestrate long-term goals and task management. The domestic chip alignment According to the company, the new model was both trained and optimized for domestic AI Application-Specific Integrated Circuit clusters, a position that is required because China has been intermittently choked off from access to Nvidia Corp.'s most powerful CUDA-based graphics processing units and chipsets. Although Nvidia chips can currently flow to China, the turbulence caused by export controls has prompted the country to seek alternatives. According to a report from Bernstein, a global equity research and brokerage firm, it was estimated in 2025 that Nvidia held around 40% of the market share in China for AI chips, roughly matched by Huawei Technologies Co., Ltd. Bernstein predicted Nvidia's market share will fall by 8% this year, giving Huawei room to grow. The model's training origin means it will run reliably and likely perform well on domestically available chips in China, while reducing dependence on Nvidia-specific software and its market dominance. The company said it was trained on ASIC "superpods," which suggests enterprise deployment within the same ecosystem and not on third-party hardware. At 1.6 trillion parameters, LongCat-2.0 will not be showing up for consumer hardware anytime soon, and it's unlikely to run on premises for most enterprise workloads. At that size, it will live in a data center or cloud environment, where it can be distributed across high-density inference clusters under management, with model parallelism. If it really is architected the way Meituan claims, then its core reasoning is portable to other hardware, but the performance optimizations will remain on domestic chips.

China's Meituan says new AI model trained on domestic chips

Chinese tech giant Meituan has unveiled LongCat-2.0, a new AI model comparable to Google's Gemini 3.1 pro. This marks a significant achievement as it's reportedly the first trillion-parameter model trained entirely on domestically developed computer chips. This development is a crucial step for China in its pursuit of AI dominance amidst US chip export restrictions, showcasing their growing self-reliance in advanced hardware for AI development. Chinese tech firm Meituan launched a new artificial intelligence model on Tuesday that it said was the first of its size to be trained using domestically developed computer chips. The country is locked in a fierce race for AI dominance with the United States, which restricts the export to China of the most cutting-edge silicon chips designed by industry leader Nvidia. In response to Washington's restrictions, imposed on national security grounds, China has been accelerating efforts to develop its own advanced chips and break away from reliance on US hardware. Meituan on Tuesday unveiled LongCat-2.0, a new large language model whose performance it says is comparable to Google's Gemini 3.1 pro, released in February. LLMs are the kind of technology that underpin chatbots and many other AI tools. LongCat-2.0 is "the industry's first trillion-parameter model to complete end-to-end training and inference on a 50,000-chip domestic compute cluster", Meituan said in a statement. It did not disclose which Chinese chipmaker's products were used in the training process. The announcement marks a milestone for China's fast-developing AI industry, because training competitive models with vast amounts of digital data requires chips packing hefty computing power. Flagship models from other Chinese AI labs, including DeepSeek and Zhipu, can run inference on chips produced by domestic companies like Huawei -- a less intensive process than training. It is believed that most domestic models were trained using Nvidia chips. Some media reports have suggested that an unknown number of advanced Nvidia AI chips were smuggled into China. Meituan's AI research team began exploring the use of domestic chips in 2023 and "has proven that we are now capable of carrying out large-scale model training on domestic compute clusters", the company said.