--- title: "Meituan open-sources LongCat-2.0, a trillion-parameter model built on 50,000 domestic chips" date: 2026-10-25 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-meituan-longcat-20 language: en --- # Meituan open-sources LongCat-2.0, a trillion-parameter model built on 50,000 domestic chips > Meituan released and open-sourced LongCat-2.0, a 1.6-trillion-parameter model trained and served entirely on a 50,000-card domestic compute cluster — a stress test for China's homegrown AI stack. A delivery platform is, under the hood, one of the hardest coding problems in the world: millions of orders, couriers, restaurants, and routes colliding in real time. Meituan's answer was to build its own trillion-parameter model and train it on Chinese-made chips — then give the weights away. ## Key takeaways - Meituan released **LongCat-2.0** and open-sourced it under an MIT license; the model has 1.6 trillion total parameters, roughly 48B average active, with a dynamic range of 33B–56B per token. - It is described as the first trillion-parameter model to complete both training and inference on a **50,000-card domestic (国产) compute cluster**, a notable vote of confidence in China's homegrown accelerators. - The architecture is purpose-built for **Agentic Coding** — understanding, generating, and executing code inside long, multi-step workflows — with a native 1M-token context. - Training used over 30 trillion tokens of multilingual and code data; Meituan reports a 1.5x model-FLOPs-utilization gain and steady-state throughput above 1 trillion tokens per day. - Inference adapters for Huawei Ascend (昇腾), Moore Threads (摩尔线程), and Muxi (沐曦) were published alongside the release, letting the trillion-parameter model run on existing domestic hardware. ## The bet: national chips, national model The headline is not the parameter count — it is where the model was trained. Running a 1.6T-parameter model through full pre-training and inference on 50,000 domestic accelerators means the team had to solve operator adaptation, communication optimization, and distributed stability from the ground up. Meituan says it cut the monthly daily-failure rate by more than 70% through fault recovery and elastic scaling, and lifted training efficiency 1.5x. If that holds up, the implication is larger than one product: it suggests China's domestic compute stack can carry frontier-scale training, not just inference. That is the kind of claim the whole industry has been waiting to see validated at trillion-parameter scale. ## Why a food-delivery company builds frontier models Meituan's business is a real-time optimization problem with a code-shaped core. Agentic Coding — models that can read a codebase, write a fix, run it, and recover from failure — maps directly onto how the company automates logistics, merchant tooling, and internal systems. Building the model in-house keeps that capability strategic rather than rented. Before the public release, a preview of LongCat-2.0 was available through OpenRouter and longcat.ai, where Meituan says it climbed into the top three globally by call volume, sitting behind only a couple of the most-used closed models — a sign that agent developers, not just benchmarks, found it useful. ## What the architecture actually does LongCat-2.0 introduces LongCat sparse attention and N-gram Embedding, aimed at making long-context handling and token-level representation cheaper and more stable. A dynamic-activation scheme lets the model flex between 33B and 56B active parameters depending on the task, so a simple request does not pay for a trillion-parameter brain. Native 1M context matters for agent work: a coding agent may need to hold hundreds of tool calls, file reads, and intermediate states in memory across a long session. That is exactly the regime where smaller models fall over. ## Open weights with the hard part included Meituan did not just drop a checkpoint. It published inference code tuned for memory- and bandwidth-constrained domestic chips, with BF16, FP8, and INT8 versions covering different deployment tiers. The explicit goal is to let owners of "stock" domestic hardware — cards already bought and sitting in data centers — actually run a trillion-parameter model instead of leaving them idle. That is a different kind of openness from "here are the weights, good luck." It is openness aimed at unlocking stranded domestic compute, which is a quietly strategic move if chip supply stays uncertain. ## Honest limitations - Core parameters (1.6T total, ~48B active, 50,000 domestic cards, MIT license) are corroborated by both Meituan's technical team and a Shanghai Securities News report, but the efficiency and stability claims are the company's own and were not independently reproduced by us. - "Domestic chips" spans several vendors with different maturity; real-world throughput on each will vary, and published numbers reflect Meituan's optimized cluster, not a typical deployment. - Top-three OpenRouter call-volume ranking is Meituan's statement about its preview; leaderboard positions shift weekly and reflect adoption, not capability alone. - Agentic Coding strength on public benchmarks does not guarantee performance on a specific private codebase; integration cost and eval design still dominate outcomes. - We did not run LongCat-2.0; this article reports the launch materials, not hands-on testing. ## What readers can do now - **Engineers on domestic hardware:** if you own Ascend, Moore Threads, or Muxi cards, the published adapters are a rare chance to benchmark a trillion-parameter model on your own cluster rather than guessing. - **Enterprises:** treat LongCat-2.0 as a concrete test case for "can we run frontier-scale AI without the usual accelerators?" — pilot it on one internal agentic-coding workflow. - **Observers:** watch whether 50,000-card domestic training becomes repeatable across labs; that, more than any single model, is the signal that matters for the supply chain. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-meituan-longcat-20 Free to quote with attribution and a link to the original.