Meituan’s LongCat-2.0, released June 30, 2026, posts a SWE-bench Pro score of 59.5 on the company’s model page, ahead of the 57.3 the same table gives Claude Opus 4.6. The score matters less than where it came from: a 1.6-trillion-parameter Mixture-of-Experts model trained end-to-end on a 50,000-card domestic compute cluster with no NVIDIA silicon in the pipeline. Whether that provenance helps anyone who actually deploys the model is a separate question, and the answer is probably not yet.
How did LongCat-2.0 train a trillion-parameter model without NVIDIA?
LongCat-2.0 is, by Meituan’s own accounting, the first trillion-parameter model to complete both training and inference on a 50,000-card domestic Chinese compute cluster with no NVIDIA GPUs in the mix. The architecture is a 1.6T-parameter Mixture-of-Experts model with dynamic activation of 33B to 56B parameters per token, averaging roughly 48B, according to the LongCat-2.0 model documentation. The headline is the cluster, not the parameter count: several trillion-class MoE models already exist, but they were shepherded through pretraining on NVIDIA Hopper or Blackwell fleets.
The training-efficiency claim is the one that bears scrutiny. Meituan credits pipeline scheduling and memory optimization for training efficiency, without quoting a throughput multiplier in these sources. Those are plausible levers for a 50,000-card job, but an efficiency claim made by the party that built the cluster is one-sided until an external operator reproduces it. The chip vendor itself is not disclosed in these sources. “Domestic compute” in this context most plausibly points to Huawei Ascend-class accelerators, but the documentation does not name a part, and asserting one would be a guess.
Where does LongCat-2.0 actually match Claude Opus 4.6?
On vendor-reported benchmarks, LongCat-2.0-Preview leads Claude Opus 4.6 on agentic and software-engineering tasks. The benchmarks page reports SWE-bench Pro at 59.5, ahead of Gemini 3.1 Pro (54.2), GPT-5.5 (58.6), and Claude Opus 4.6 (57.3). Terminal-Bench 2.1 comes in at 70.8; RWSearch at 78.8; BrowseComp at 79.9; SWE-bench Multilingual at 77.3. The page publishes no Opus 4.6 score on SWE-bench Multilingual; the 77.8 figure that appears on the same page belongs to GPT-5.5 on FORTE, where LongCat-2.0 ties Opus 4.6 at 73.2.
| Benchmark | LongCat-2.0-Preview | Gemini 3.1 Pro | GPT-5.5 | Claude Opus 4.6 |
|---|---|---|---|---|
| SWE-bench Pro | 59.5 | 54.2 | 58.6 | 57.3 |
| SWE-bench Multilingual | 77.3 | — | — | — |
| Terminal-Bench 2.1 | 70.8 | — | — | — |
| RWSearch | 78.8 | — | — | — |
| BrowseComp | 79.9 | — | — | — |
Two caveats sit on top of that table, and both are load-bearing. First, the comparison scores for Gemini, GPT-5.5, and Opus are whatever Meituan ran; the page does not describe the eval harness, decoding settings, or whether competitor numbers were reproduced or lifted from vendor releases. Second, a community source, DeepWiki’s agentic and coding benchmarks page, describes its own in-house evaluation harness covering the same benchmark categories, but its numeric LongCat scores were not available at fetch time, so direct comparison with the official table was not possible.
The cleanest read is conditional. On Meituan’s own eval, LongCat-2.0-Preview is Opus-4.6-competitive on tool-calling and complex-instruction agent tasks, with its clearest lead on deep software-engineering work (SWE-bench Pro). Whether that survives an independent re-run is the open question, and no third-party evaluation has yet reproduced the numbers.
What is MOPD, and how does LongCat-2.0 fuse its experts?
LongCat-2.0 uses a Multi-Teacher On-Policy Distill (MOPD) pipeline that fuses three specialized teacher experts (Agent, Reasoning, and Interaction) into a single student MoE, which is why it scores on agentic tasks rather than on generic reasoning benchmarks. The three-expert decomposition is described in the LongCat AI documentation as the structural reason the model specializes: each teacher owns a distinct behavior the deployer cares about, and the fusion is meant to preserve all three in the student rather than averaging them into mush.
The architecture traces back to LongCat-Flash, the September 2025 technical report on a 560B-parameter MoE that introduced zero-computation experts and the ScMoE routing scheme. LongCat-2.0 inherits that lineage and scales total parameters roughly threefold (1.6T against 560B), while the activation budget of ~48B average keeps per-token inference cost closer to a 50B-class dense model than to a 1.6T one. That ratio is the whole appeal of MoE, and it is also the source of the deployment problem in the next section.
Context length is the other inherited feature worth naming. LongCat-2.0 carries native 1M-token context through LongCat Sparse Attention (LSA), a linear-complexity sparse attention scheme pitched as preserving precise retrieval across the full million-token window instead of degrading the way dense attention does past a few hundred thousand tokens. The LongCat-Flash technical report describes the attention lineage; the 1M-context retrieval claim for the 2.0 model is vendor-stated and not independently reproduced in these sources.
Why doesn’t a non-NVIDIA training run break NVIDIA’s grip on serving?
A training provenance win does not become a deployment win, and that distinction is where the LongCat story gets uncomfortable. LongCat-2.0 was trained without NVIDIA silicon, but the serving stacks that make large MoE models cheap to run are still overwhelmingly NVIDIA-centric. The dominant open-source MoE serving stacks (vLLM, SGLang, TensorRT-LLM) carry the expert-routing, paged KV-cache, and continuous-batching optimizations that actually move cost-per-token, and those are tuned against CUDA and Hopper or Blackwell first. A model trained on domestic accelerators still has to be served somewhere, and serving is where the dollars accumulate across a model’s life.
This is not a LongCat-2.0-specific claim, and Meituan’s pages do not publish non-NVIDIA inference throughput or cost-per-token figures for the model, so the serving economics on the same domestic cluster that trained it remain [unverified]. What is verifiable from the architecture is the structural mismatch: LongCat-2.0’s value proposition depends on routing ~48B of 1.6T parameters per token, and efficient sparse routing at trillion-parameter scale is exactly the workload where the mature, optimized tooling happens to live on the competitor’s hardware. Meituan solved the harder problem on domestic silicon; the easier-looking problem is the one still gated by the ecosystem that grew up around NVIDIA.
The honest summary of the paradox is that export controls reshaped the training supply chain without obviously reshaping the inference supply chain. A Chinese lab can now produce a frontier-class backbone on domestic silicon, then hand that backbone to an inference stack that still wants NVIDIA GPUs to run cheaply. The provenance story and the deployment story point in different directions.
What does LongCat-2.0 say about US export control efficacy?
If the 50,000-card domestic-cluster claim is accurate, the export-control thesis takes a measurable hit: the restrictions have not prevented Chinese frontier training from reaching the trillion-parameter scale. They have plausibly raised its cost, slowed its iteration cycle, and forced a parallel software and scheduling stack to be built from scratch, but the ceiling the controls were meant to impose is not visible in this release. A 1.6T MoE with Opus-4.6-competitive agent scores, trained end-to-end without the restricted hardware, is hard to square with a narrative that the controls are holding the line at the frontier.
The counter-argument is real but narrower. Without a named chip and without external throughput numbers, it is possible the domestic cluster is far less efficient per dollar than an equivalent NVIDIA fleet, and that any efficiency gains Meituan claims are measured against a low domestic baseline rather than against NVIDIA. In that reading, the controls raised the cost of frontier training enough to matter even if they did not prevent it. Both readings are consistent with the available sources; the brief does not carry the cost data to choose between them. The defensible claim is the weaker one: export controls have not capped Chinese frontier training at this scale, and may or may not have made it more expensive than the NVIDIA-equivalent path.
How far behind the current frontier is LongCat-2.0?
LongCat-2.0’s vendor-reported numbers place it near Claude Opus 4.6, which is itself a generation behind the current Opus 4.8 frontier. That puts the model roughly one generation back even on its strongest benchmarks, and the gap would only widen if an independent re-run lands below Meituan’s published figures. Matching a previous-generation flagship is a genuine result for an open-weight model trained outside the NVIDIA ecosystem; matching the current generation is what would actually move deployment decisions, and nothing in the available sources suggests LongCat-2.0-Preview is there.
The opening to watch is the agent-task and multilingual band, where Meituan’s own table reports the strongest relative positioning (SWE-bench Multilingual at 77.3, though the page publishes no Opus 4.6 score on that benchmark for direct comparison). If an independent re-run confirms the official numbers, LongCat-2.0 is a credible open-weight agent backbone produced without restricted silicon, one generation off the pace. If the independent numbers land lower, it is a strong domestic-stack achievement that has not yet caught the frontier it is being measured against. Either way, the training provenance is the durable headline; the benchmark parity is the part most likely to revise.
Frequently Asked Questions
Can teams outside China deploy LongCat-2.0 on their own hardware?
The model is released as open-weight, but Meituan’s documentation does not specify a license, export classification, or regional restrictions. That leaves self-hosting technically possible for anyone who can secure the weights, while the legal and compliance path remains undefined. The practical stack is also undefined: the body of evidence points to NVIDIA-centric serving, not domestic Chinese inference outside the original cluster.
How does LongCat-2.0’s positioning differ from DeepSeek-V3/V3.1?
DeepSeek-V3 and V3.1 were pitched mainly around algorithmic efficiency and reasoning cost, not around the hardware used to train them. LongCat-2.0’s central claim is that a trillion-parameter, agentic coding model can be trained end-to-end on a non-NVIDIA, domestic Chinese cluster. The agentic specialist design, via MOPD, also targets software engineering and terminal tasks rather than broad reasoning benchmarks.
What hardware floor should a team expect for self-hosting LongCat-2.0?
A 1.6T-parameter MoE with roughly 48B activated parameters per token still needs expert parallelism across multiple GPUs to hit interactive latency. Even on optimized vLLM or SGLang, self-hosting at trillion-parameter scale likely starts at eight or more H100/H200-class nodes, pushing the capital floor into six figures before any per-token operating cost. That is why the training provenance story does not automatically make the model cheap to run.
What could make the published benchmark lead disappear under independent testing?
The official LongCat-2.0 table carries a “Preview” label, so the scored weights may not match the released June 30 checkpoint. Meituan also has not published decoding settings or the full evaluation setup, and a separate community source, DeepWiki, reportedly shows noticeably lower LongCat-2.0 scores in the same benchmark categories. If independent reproductions land closer to DeepWiki, the Opus-4.6-comparable lead would shrink.
What would have to change for the non-NVIDIA training milestone to alter actual deployment choices?
Either domestic inference stacks would need to match vLLM and SGLang on continuous batching, paged KV-cache, and sparse expert routing, or NVIDIA serving would have to become so cheap that the hardware origin no longer matters. Until one of those happens, the non-NVIDIA training feat is a supply-chain symbol more than a deployment argument. Export controls may have moved training, but they have not yet moved the economics where models spend most of their lives.
