For years, the working assumption in tech policy circles has been blunt: no NVIDIA chips, no frontier AI. The export con

For years, the working assumption in tech policy circles has been blunt: no NVIDIA chips, no frontier AI. The export controls on high-end GPUs like the H100 and the newer Blackwell line were designed precisely around that bottleneck. Cut off access to the best silicon, the theory went, and you slow a country’s ability to train the largest, most capable models. Meituan just called that bluff.

The Chinese tech giant, best known for dominating food delivery and local services, has released LongCat 2.0. It is a 1.6-trillion-parameter model built on a Mixture-of-Experts architecture. The kicker? The team trained it entirely on domestic Chinese chips. No NVIDIA H100s. No Blackwell GPUs. And rather than hiding it behind an API, Meituan has dropped the weights, training code, and full data pipeline on GitHub for anyone to inspect.

What LongCat 2.0 Actually Brings

Let’s look at the hardware-agnostic facts first. At 1.6 trillion parameters, LongCat 2.0 sits among the largest open-weight models ever released. Parameters are not the only measure of quality, but at this scale, they signal serious engineering ambition. Because it uses a Mixture-of-Experts design, only a subset of those parameters activates for any given task. That keeps inference costs from ballooning out of control while still allowing the model to store an enormous range of knowledge and reasoning patterns.

The context window hits one million tokens. That matches the high-end offerings from OpenAI and Anthropic, and it changes what the model can practically do. With a window that large, you can feed it entire legal contracts, months of chat logs, or vast code repositories in a single prompt. For developers building applications that require memory across long documents, this is not a marginal improvement. It is a functional necessity.

Then there is the openness. Meituan did not release a sanitized demo or a restricted API. The model weights are available, the training code is public, and so is the pipeline. That matters for researchers who want to reproduce results, for enterprises that need to audit behavior, and for engineers who need to fine-tune the model on proprietary data without sending anything to a third-party server.

The Hardware Story Everyone Missed

The headline here is not just the parameter count. It is the silicon underneath. LongCat 2.0 was trained on domestic accelerators, specifically chips like the Huawei Ascend series. That is a radically different proposition from slapping an existing framework onto an NVIDIA cluster and hitting run.

Training a trillion-parameter model requires solving distributed computing problems at the extreme edge. Memory bandwidth, inter-chip communication, and floating-point performance all have to be managed with ferocious precision. NVIDIA’s advantage has never been just the raw GPUs. It is the CUDA ecosystem, the optimized kernels, and the collective knowledge of how to squeeze performance out of that stack. Building a comparable model on Ascend hardware means Meituan’s engineers had to do the grueling work of adapting training frameworks, rewriting low-level operations, and debugging distributed training runs on a fundamentally different architecture.

That they succeeded suggests something larger than one model. It shows that Chinese hardware-software co-design is maturing to the point where the absence of Western chips is no longer a hard stop. It is a constraint, and an expensive one, but not an impossible barrier.

Where It Actually Performs

LongCat 2.0 scores strongly in Chinese reasoning, mathematics, and long-context retrieval. Those are specific, high-value benchmarks. Mathematical reasoning tests a model’s ability to handle logic and symbolic manipulation. Long-context retrieval tests whether the model can find a needle of information in a million-token haystack without losing track. Passing both is the difference between a model that sounds smart and one that can actually do work.

Meituan se positionne comme une alternative directe à Llama, DeepSeek et Qwen. Pour les tâches en langue chinoise, la concurrence est particulièrement vive. Les modèles entraînés principalement sur des données internet en anglais trébuchent souvent sur les références au chinois classique, le langage réglementaire local, la terminologie financière utilisée sur les marchés continentaux et le jargon informel qui domine les réseaux sociaux chinois. Un modèle conçu avec de profondes compétences linguistiques en chinois, alimenté par une entreprise dont le cœur de métier repose sur le comportement des consommateurs chinois, possède un avantage structurel dans ces scénarios.

Cela importe pour bien plus que les chatbots. Les entreprises de Legaltech analysant le droit des contrats de la RPC, les chercheurs décortiquant des textes historiques, les banques traitant de la documentation de prêt en mandarin et les plateformes de service client gérant les dialectes régionaux ont tous besoin de modèles qui comprennent les nuances plutôt que de traduire à travers un prisme centré sur l'anglais.

L'arme cachée de Meituan : l'échelle et les données

Meituan n'est pas un laboratoire de recherche avec une application de livraison en complément. C'est un géant opérationnel qui coordonne chaque jour des millions de livreurs, de restaurants et de commerçants à travers la Chine. Cette échelle génère un torrent de données linguistiques du monde réel. Des plaintes de clients acheminées vers des chatbots. Des descriptions de restaurants mêlant l'argot régional aux détails formels d'enregistrement commercial. Des instructions d'itinéraire mélangeant systèmes d'adresses et points de repère informels. Des tickets de service marchand traitant de la fiscalité, des réglementations d'hygiène et des ordonnances locales.

Ces données sont désordonnées, contextuelles et profondément locales d'une manière que les données génériques de crawl web ne pourront jamais reproduire. L'intégration de ces données dans LongCat 2.0 offre au modèle un ancrage pratique qui fait défaut aux cycles d'entraînement purement académiques. C'est une chose d'entraîner un modèle sur des textes Wikipedia so