LongCat-2.0

2026-06-29 · Publisher : Meituan · Domain : Artificial Intelligence

Meituan announces LongCat-2.0, a mixture-of-experts model of 1.6 trillion parameters with about 48 billion active, under the MIT license, with a one-million-token context, trained on Chinese ASICs.

Meituan announces LongCat-2.0, a mixture-of-experts model of 1.6 trillion parameters with about 48 billion active, under the MIT license, with a one-million-token context, trained on Chinese ASICs. ### Context LongCat-2.0 succeeds LongCat-Flash (September 2025) and its Thinking, Omni and Lite variants. It targets the class of frontier closed models of summer 2026: Gemini 3.1 Pro, GPT-5.5 and Claude Opus 4.6 to 4.8, which the model card uses for comparison. The announcement is dated June 29; the weights have been on Hugging Face since early July. ### What the model brings - LongCat Sparse Attention (LSA): three improvements to the indexer of DeepSeek Sparse Attention (memory-flow-adapted indexing, indexing shared across layers, hierarchical indexing from coarse to fine) to make long context less costly. - N-gram embeddings: 135 billion parameters of n-gram embeddings, inherited from LongCat-Flash-Lite, to increase capacity along a sparse dimension other than the experts. - Training on AI ASICs: the full training and the large-scale deployment took place on superpods of ASIC accelerators, with no rollbacks and no unrecoverable loss spikes, according to the model card; the press reports more than 50,000 Chinese accelerators and no NVIDIA or AMD GPUs. - 1-million-token context: hundreds of billions of tokens of data at 1 million tokens were used for the long training. - Integration into agent harnesses (Claude Code, OpenClaw, Hermes) for repository-level code and the execution of automated tasks. ### Architecture and specifications - Architecture: MoE of 1.6 trillion parameters in total, about 48 billion active per token, 3-step multi-token prediction module. - Training: more than 35 trillion tokens (according to the model card). Context: 1 million tokens. Modalities: text. - License: MIT on the model weights. - Weights: meituan-longcat/LongCat-2.0 on Hugging Face and ModelScope, with an FP8 variant according to the press. GPU and NPU deployment via SGLang. Knowledge cutoff: not disclosed. ### Cost and access According to VentureBeat, the API is offered at a promotional price of 0.30 dollars per million input tokens and 1.20 dollars for output, then at a standard price of 0.75 and 2.95 dollars, with free cache. Aggregators recommend 16 H20 GPUs for deployment (secondary source, not confirmed by the model card). | Item | Price (USD per million tokens) | | --- | --- | | Input (promotional) | 0.30 | | Output (promotional) | 1.20 | | Input (standard) | 0.75 | | Output (standard) | 2.95 | | Cache hit | Free | ### Benchmarks | Benchmark | LongCat-2.0 | Gemini 3.1 Pro | GPT-5.5 | Claude Opus 4.7 | Claude Opus 4.8 | | --- | --- | --- | --- | --- | --- | | Terminal-Bench 2.1 | 70.8 | 70.7* | 73.8* | 71.7* | 78.9* | | SWE-bench Pro | 59.5 | 54.2* | 58.6* | 64.3* | 69.2* | | SWE-bench Multilingual | 77.3 | 76.9* | - | 80.5* | 84.8* | | BrowseComp | 79.9 | 85.9* | 84.4* | 79.3* | 84.3* | | GPQA-Diamond | 88.9 | 94.3* | 93.6* | 94.2* | 92.4* | | IFEval | 90.0 | 96.1 | 95.0 | 88.7 | 86.0 | | IMO-AnswerBench | 81.8 | 90.0 | 79.5 | 81.8 | 75.3 | Source: Hugging Face model card of LongCat-2.0. LongCat-2.0 scores measured internally by Meituan; *: value quoted from the official report of the model concerned, otherwise measured by Meituan under a unified harness. ### Versus closed models On agent code, LongCat-2.0 falls within the range of the frontier closed models cited: 59.5 on SWE-bench Pro, ahead of Gemini 3.1 Pro (54.2) and GPT-5.5 (58.6), but behind Claude Opus 4.7 (64.3) and Opus 4.8 (69.2); on Terminal-Bench 2.1, 70.8, level with Gemini 3.1 Pro (70.7) and 1 point behind Opus 4.7 (71.7). The gap is clearer on knowledge (GPQA-Diamond 88.9 versus 92.4 to 94.3 for the four closed models) and on BrowseComp (79.9 versus 79.3 to 85.9, in the same range as Opus 4.7). On IMO-AnswerBench, it is ahead of GPT-5.5 and Opus 4.8 (81.8 versus 79.5 and 75.3). The model card presents Opus 4.6 to 4.8 as references, and VentureBeat writes that the model remains behind premium systems such as Claude Opus 4.8 on broader benchmarks. On pricing, 2.95 dollars per million output tokens at the standard rate, according to VentureBeat, is the only reference point; closed-model pricing is not recorded here. For an open-versus-closed choice, the appeal is twofold: an open-weights model of the trillion-parameter class under MIT, and training carried out without US GPUs, an argument for hardware sovereignty. ### Key takeaways LongCat-2.0 takes Meituan from 560 billion to 1.6 trillion parameters, with a context of one million tokens and an MIT license. According to the vendor's figures, it is close to GPT-5.5 and Gemini 3.1 Pro on agent code but behind Claude Opus 4.7 and 4.8.

Tags : LLM, Open Source, Meituan, LongCat, MIT, MoE, 1,6 T