Meituan, best known for food delivery and local services, released LongCat-2.0 on June 30, 2026: a 1.6-trillion-parameter Mixture-of-Experts model with roughly 48 billion parameters active per token, trained and served end-to-end on a 50,000-unit cluster of domestic Chinese accelerators. The weights ship under an MIT license. Every figure is vendor-reported and unverified by third parties as of 2026-07-08, but if the claims hold, the consequence is structural: a consumer-internet operator built a frontier-scale model on hardware the US export regime was meant to deny it.
What is LongCat-2.0?
LongCat-2.0 is a 1.6-trillion-parameter MoE model that activates approximately 48 billion parameters per token across a dynamic range of 33B to 56B, trained on over 35 trillion tokens with native 1-million-token context, according to Meituan’s release blog. The headline numbers put it in the same weight class as the largest Chinese open weights from DeepSeek and Zhipu, but the provenance is what makes it unusual. Meituan is a food-delivery, hotel-booking, and local-services platform, not a dedicated AI laboratory. A company whose core business is dispatching riders to restaurants full-pretraining a 1.6-trillion-parameter model is the part that warrants attention, not the parameter count on its own.
Was LongCat-2.0 really trained without US GPUs?
Meituan claims LongCat-2.0 is the first trillion-parameter model to complete full end-to-end training and inference on a 50,000-unit domestic compute cluster, with no chip vendor officially confirmed as of 2026-07-08 (model page). Huawei’s Ascend line is the widely assumed supplier, but Meituan has not named a vendor, and the company’s own wording distinguishes “AI ASIC” from conventional GPU, which suggests the accelerator architecture departs from the NVIDIA-style design most training stacks assume.
That claim is the load-bearing one for the whole release. The US export-control framework is premised on the idea that denying access to advanced NVIDIA and AMD GPUs confines frontier-scale training to a small set of actors who either have domestic accelerator programs mature enough to substitute or can move enough restricted hardware to matter. A food-delivery company clearing the trillion-parameter bar on a 50,000-card domestic cluster, if accurate, means the substitution path works at frontier scale outside the national-champion labs.
Meituan reports that it began domestic-compute exploration in 2023 and scaled from thousands of cards to 50,000, achieving a better-than-70% reduction in daily fault rate, a 1.5x improvement in model FLOPs utilization (MFU), and steady-state throughput exceeding 1 trillion tokens per day, all vendor-reported. These are operational metrics for the training cluster, not model-quality metrics. The fault-rate and MFU numbers, if genuine, are the harder figures to fake, because they describe an engineering grind sustained over years rather than a single benchmark run that can be cherry-picked.
How does LongCat-2.0’s architecture work?
The architecture rests on three components Meituan names LongCat Sparse Attention, Zero-Computation Experts, and MOPD multi-expert fusion, according to the project’s GitHub repository. LongCat Sparse Attention provides linear-complexity attention the model uses to support its native 1-million-token context without the quadratic memory cost standard attention would impose at that length. Zero-Computation Experts handle token-level dynamic activation, the mechanism behind the 33B-to-56B activation range: rather than a fixed expert set firing per token, the activated parameter count varies with the input. MOPD fuses separate Agent, Reasoning, and Interaction expert groups, which is consistent with Meituan positioning the model for tool-using workflows rather than open-ended chat.
The repo also describes a 135-billion-parameter N-gram Embedding component that accounts for under 10% of the total parameter budget. The stated rationale is parameter-utilization efficiency: by expanding parameters in sparse dimensions orthogonal to the MoE routing, the model grows capacity without proportionally increasing per-token activation cost. Whether this holds up under independent inspection is an open question, but the design choice is internally coherent. Most large MoE models spend their parameter budget on expert weights; adding a large embedding block that does not activate per token is a recognizable strategy for raising capacity at sublinear inference cost.
How does LongCat-2.0 perform on coding and agentic benchmarks?
On Meituan’s in-house suite, LongCat-2.0 reports SWE-bench Pro 59.5, SWE-bench Multilingual 77.3, Terminal-Bench 2.1 70.8, FORTE 73.2, RWSearch 78.8, and BrowseComp 79.9, all measured under the company’s unified harness and none independently reproduced (benchmark page). Meituan’s own comparison places LongCat-2.0’s SWE-bench Pro score ahead of GPT-5.5 at 58.6, Claude Opus 4.6 at 57.3, and Gemini 3.1 Pro at 54.2.
The benchmark selection is itself a signal. SWE-bench Pro, Terminal-Bench, FORTE, RWSearch, and BrowseComp are all agentic or coding-oriented evaluations, not general-knowledge suites like MMLU. That aligns with a community review describing LongCat-2.0 as an agentic coding model integrated into the Claude Code, OpenClaw, and Hermes harnesses for repository-level edits, automated task execution, and tool use. Meituan is not claiming a general-purpose assistant. It is claiming a coding and agent model, which is a narrower and more defensible claim if the numbers turn out to be soft.
Can you self-host LongCat-2.0?
The weights are released under an MIT license and are available on longcat.ai and OpenRouter, where LongCat-2.0 reportedly ranked among the top three models globally by call volume before the public announcement, according to an industry analysis. MIT is the most permissive license in common use for model weights, with no commercial-use restrictions and no copyleft obligations, which removes the licensing friction that has slowed enterprise adoption of some Chinese open weights.
For procurement teams that currently route inference to US-hosted APIs, the practical question is whether a self-hostable 1.6-trillion-parameter Chinese model is a viable substitute. A model this large is not trivial to serve: 1.6 trillion parameters in fp16 is on the order of 3 terabytes of weights to hold in memory, and even with 48 billion active per token, the full expert pool must reside on-device. Self-hosting is realistic only for operators with a substantial accelerator fleet, which is exactly the constituency Meituan’s domestic-cluster story is aimed at. Operators without that fleet get the API on OpenRouter, not the weights.
What is unverified, and what should you watch for?
Every quantitative claim in this article traces back to Meituan or to coverage that repeats Meituan’s figures, and as of 2026-07-08 none of it has been independently benchmarked. The specific gaps to track: which domestic chipmaker supplied the 50,000 accelerators, and whether the ASIC-versus-GPU wording reflects a genuinely different architecture or is marketing framing; whether any third-party lab reproduces the SWE-bench Pro, Terminal-Bench, and agentic scores under a harness Meituan does not control; and whether the 1-trillion-tokens-per-day throughput and 1.5x MFU figures hold up to outside scrutiny of the training run.
The strategic claim is the one most resistant to verification and the one that matters most. If a food-delivery company can full-pretrain a 1.6-trillion-parameter model on domestic accelerators and ship it under MIT, then the binding constraint on sovereign-compute LLM strategies is not access to US silicon. It is having a 50,000-card fleet and the engineering patience to make it reliable. That is a lower bar than the export-control framing assumes, and it applies to any operator with a domestic accelerator fleet, not just China’s designated national champions. Watch for independent evals and a named chip vendor. Until then, this is a vendor-reported proof of concept for a thesis the policy debate has not fully priced in.
Frequently Asked Questions
Who can realistically run LongCat-2.0 on their own hardware?
Self-hosting requires holding roughly 3 terabytes of weights in memory for fp16, plus enough accelerator interconnect and context-length kernels to serve a 1-million-token window. Most teams will use the OpenRouter API instead; only operators with a large fleet and MoE-serving expertise should plan an on-prem deployment.
How does LongCat-2.0 compare to DeepSeek-V3?
DeepSeek-V3 is also a large MoE, but its reported 671 billion total and 37 billion active parameters are smaller than LongCat-2.0’s 1.6 trillion total and 48 billion active. The bigger difference is hardware provenance: DeepSeek trained and serves on NVIDIA H800 and H20-class GPUs, while Meituan claims an all-domestic Chinese ASIC cluster of 50,000 accelerators.
What changes if a team switches from a US-hosted API to LongCat-2.0?
Moving inference in-house means leaving the CUDA ecosystem if you run on the domestic Ascend stack Meituan implies, with CANN compilers and HCCL networking instead of NCCL. You also need an inference engine that supports LongCat Sparse Attention and dynamic expert activation; stock vLLM or SGLang may not cover those kernels on day one.
What could make the published benchmarks look better than real performance?
All scores were measured by Meituan’s own evaluation setup, and SWE-bench variants are notoriously sensitive to prompt wording and tool configuration. Training data can also include public benchmark solutions, so high scores may reflect memorization rather than general coding ability until an outside lab reproduces them under a different test rig.
What would prove or disprove the domestic-training claim?
The decisive evidence is independent verification of the training run: a named chip vendor, reproducible MFU and fault-rate curves, and third-party throughput logs showing sustained delivery above one trillion tokens per day. If the hardware turns out to be smuggled NVIDIA or AMD cards, the export-control story collapses even if the model itself is solid.
