Meituan has published a model that is open in the legal sense, open in the inspectable-artifact sense, and emphatically not small. On July 6, the Chinese local-services company released LongCat-2.0's weights under the MIT License and published inference code. The main Hugging Face checkpoint is divided into 194 safetensors shards. Adding their listed sizes gives 3.55 terabytes; the separate FP8 and INT8 repositories each remain just over 2 terabytes.[2][3]

That scale is the most useful way to read the release. LongCat-2.0 is not a laptop model wearing a frontier label. It is a cluster-native artifact whose openness lets outsiders inspect, serve, adapt, and test it if they have serious infrastructure. The release therefore separates two ideas that are too often collapsed: access to weights and ease of deployment.

It also turns Meituan's June 30 announcement into something more concrete. The first statement said the model had been trained from scratch on a cluster of more than 50,000 domestically produced accelerators and would be opened later. Six days afterward, model files, a public repository, multiple precision variants, a GPU route, and an NPU implementation appeared.[1][2] That sequence matters. A hardware-sovereignty claim is still a vendor claim; a published software package gives engineers something they can actually interrogate.

The release arrives in two layers

The model layer is immense but sparse. Meituan reports 1.6 trillion total parameters, with about 48 billion activated per token on average and a dynamic range of 33 billion to 56 billion. The June announcement said training used more than 30 trillion tokens, including Chinese, English, other languages, and code; the current model card gives the more specific figure of more than 35 trillion, with hundreds of billions of tokens devoted to one-million-token context training.[1][3]

Sparse activation lowers the computation performed for a token; it does not make the full parameter set disappear. All those expert weights still have to be stored, distributed, loaded, and made reachable when the router calls them. That is why the public checkpoint can occupy several terabytes even though the active path resembles a much smaller model. LongCat-2.0's release package makes the distinction physical.

The hardware layer is Meituan's larger claim. Its June account says three years of work on domestic accelerators improved fault handling, elastic replacement, deterministic operators, communication, memory use, and pipeline scheduling. The company reports a greater-than-70-percent reduction in its average daily fault rate, a 1.5-fold improvement in model FLOP utilization, and steady-state training throughput above one trillion tokens per day.[1] Reuters independently reported the announcement but correctly preserved the attribution: Meituan made the 50,000-chip claim and did not identify the chipmaker used for training.[5]

That missing name is important. The open NPU serving path targets Huawei Atlas A2 hardware, but it does not establish that the 50,000-accelerator training run used the same hardware. Inference documentation and training provenance are different evidence.

The architecture is designed around movement, not just arithmetic

Three choices explain why LongCat-2.0 is interesting beyond its parameter count.

First, LongCat Sparse Attention tries to make long-context retrieval more compatible with the memory system underneath it. Streaming-aware indexing combines contiguous reads with dynamic token selection; cross-layer indexing reuses one selection pass across adjacent layers; hierarchical indexing narrows the search from blocks to individual tokens. Meituan says this avoids the quadratic scoring bottleneck of the indexer it builds on and extends through the model's three-step multi-token-prediction path.[3]

Second, the model adds 135 billion N-gram embedding parameters. Meituan's argument is that, once the mixture-of-experts layer is already extremely sparse, adding another bank of experts produces less value than expanding token-level representation along a different sparse dimension.[2][3] This is not evidence that N-gram embeddings will generalize better on every workload. It is a visible scaling bet: spend more parameters on recurring local token patterns instead of making an already huge expert pool still wider.

Third, the model uses zero-compute experts and dynamic activation. Simpler tokens can follow a lighter route while harder ones receive more active capacity.[1][2] That makes the 33-to-56-billion range more informative than the average alone. Serving performance will depend on the distribution of routes across real traffic, not merely on a single advertised active-parameter number.

All three choices share a systems theme. Sparse models win only when theoretical savings survive memory reads, interconnect traffic, expert imbalance, and scheduling. LongCat-2.0's design is trying to make sparsity legible to the machine that must serve it.

The NPU recipe is the reality check

The strongest document in the package may be the least glamorous: an NPU deployment example in Meituan's SGLang-FluentLLM repository. It specifies an Atlas A2 cluster with 192 cards across 12 machines, dividing 64 cards for prefill and 128 for decode. The recipe names Linux x86-64, PyTorch 2.6, CANN 8.5-era software, custom operators, model paths, launch scripts, and an OpenAI-compatible request used to confirm that the service is running.[4]

This is useful disclosure because it prevents “NPU support” from floating as a logo on a model card. There is a real topology and a real software stack. It also establishes the adoption boundary. A 192-card reference deployment is not a turnkey path for a small team, and the repository does not yet provide a matched throughput, latency, energy, or cost comparison against a GPU configuration.[4]

The result is a more honest definition of open. A university lab, cloud provider, accelerator vendor, or very large enterprise can audit the model and reproduce parts of the serving path. Most developers will consume it through a hosted endpoint. That does not make the weight release cosmetic: public artifacts allow independent evaluation, derivative work, security inspection, and alternative hosting. It does mean that “downloadable” should not be confused with “democratized.” Nor is this a reproducible training release: the cited package does not include the training data, training code, or a recipe capable of recreating the 50,000-chip run.[2][3]

The benchmark table is a claim ledger, not a verdict

Meituan reports 70.8 on Terminal-Bench 2.1, 59.5 on SWE-bench Pro, and 77.3 on SWE-bench Multilingual. Those are relevant choices for a model pitched at coding agents, but the model card says that results were measured in-house unless marked as taken from another model's official report.[3] The table is therefore vendor evidence assembled from a mixture of Meituan runs and published comparison numbers, not one independent, blind evaluation.

The one-million-token claim needs a similar boundary. The release establishes architecture, training exposure, and a stated capacity, but the published config.json declares max_position_embeddings as 262,144. The cited documentation does not explain the serving override or otherwise reconcile that checkpoint default with the advertised one-million-token window. It also does not publish a length-by-length curve showing retrieval, reasoning, and agent performance as context grows, or the memory and latency cost across representative repositories.[1][3] Context capacity is an interface limit; useful long-context behavior is an evaluation result.

That leaves three different judgments, and the release is strongest when they remain separate:

  1. Artifact openness is concrete. The weights, license, precision variants, model card, and serving code are public.[2][3][4]
  2. Capability evidence is directional. The coding and agent results are substantial enough to justify outside testing, but they are still mainly vendor-reported.[3]
  3. Domestic-hardware economics are unresolved. Meituan reports a notable full-stack training achievement, while chip identity, matched cost, power, utilization, and reproducible training details remain undisclosed.[1][5]

The next meaningful evidence is therefore specific: independent agent runs with published harness settings; long-context curves rather than one endpoint number; throughput and tail-latency results for the public 192-card NPU topology; and a clearer bill of materials for the training run. If those arrive, LongCat-2.0 will be evidence that China's domestic compute stack can support not just a spectacular training event but a maintainable open-model service. If they do not, the release still matters—but principally as an unusually inspectable cluster artifact.

Sources

  1. Meituan, “LongCat-2.0 released: trillion-parameter model trained entirely on domestic chips” (June 30, 2026)—official launch statement, training-cluster claim, architecture, training data, stability and throughput figures.
  2. Meituan, “LongCat-2.0 open-sourced with domestic-accelerator inference code” (July 6, 2026)—official opening announcement, precision variants, architecture changes, deployment strategy and artifact links.
  3. Meituan LongCat Team, LongCat-2.0 model card and public artifact repository—MIT license, model design and configuration, benchmark provenance, checkpoint shards, deployment routes and usage boundaries.
  4. Meituan LongCat Team, SGLang-FluentLLM NPU deployment example—Atlas A2 192-card topology, prefill/decode split, CANN environment, custom operators and launch procedure.
  5. Ethan Wang and Eduardo Baptista, Reuters, “China's Meituan says new AI model trained on domestic chips” (June 30, 2026)—independent reporting on the announcement, company context, attribution limits and undisclosed chip supplier.
  6. Channel News Asia / AFP, “China's Meituan says new AI model trained on domestic chips” (June 30, 2026)—reporting context and source page for Greg Baker's April 27, 2021 photograph of Meituan riders in Beijing.