The LongCat-2.0 model card is the Hugging Face repository page through which Meituan's LongCat team released the weights of LongCat-2.0 on June 30, 2026 (Source: https://huggingface.co/meituan-longcat/LongCat-2.0). It is the only developer-authored document with numeric evaluation results for the model located as of September 2026; the card links a tech blog rather than a technical report, and no arXiv report for LongCat-2.0 was found (Meituan's earlier LongCat-Flash report is arXiv 2509.01322). The card covers the model's architecture in outline, a benchmark table, deployment instructions for GPU and NPU serving with vLLM and SGLang, the chat template with a thinking-mode switch, and the licence.
Model
The card describes LongCat-2.0 as a mixture-of-experts language model with 1.6 trillion total parameters and about 48 billion activated per token, pre-trained on more than 35 trillion tokens over "millions of accelerator-days" with "no rollbacks or irrecoverable loss spikes", and states that both training and large-scale deployment run "entirely on AI ASIC superpods"; the accelerator vendor is not named. Two architectural features are described. LongCat Sparse Attention (LSA) modifies the indexer of DeepSeek's sparse attention with streaming-aware indexing, cross-layer indexing (one indexing pass shared by several layers, enabled by cross-layer distillation) and hierarchical coarse-to-fine indexing, and extends to a three-step multi-token-prediction module for speculative decoding. N-gram Embedding, inherited from LongCat-Flash-Lite, adds 135B parameters in sparse dimensions orthogonal to the MoE. The model was trained on "hundreds of billions of tokens" of 1M-context data and is described as integrated with agent harnesses including Claude Code, OpenClaw and Hermes. Weights are released under the MIT licence, which the card says grants no rights to Meituan trademarks or patents.
Evaluation
The card's single table compares LongCat-2.0 with Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.6, Claude Opus 4.7 and Claude Opus 4.8 on Terminal-Bench 2.1, SWE-bench Pro and SWE-bench Multilingual (code agent); FORTE, BrowseComp and RWSearch (general agent); and IFEval, Writing Bench, IMO-AnswerBench and GPQA-diamond (foundational). Unmarked comparison scores were "measured in-house under a unified harness"; scores marked with an asterisk are "cited from the model's official report". LongCat-2.0 is reported at 70.8 on Terminal-Bench 2.1, 59.5 on SWE-bench Pro, 77.3 on SWE-bench Multilingual, 79.9 on BrowseComp and 88.9 on GPQA-diamond. In the table it leads no row outright: Claude Opus 4.8 is highest on the three code-agent benchmarks, GPT-5.5 on FORTE, RWSearch and Writing Bench, Gemini 3.1 Pro on BrowseComp, IFEval, IMO-AnswerBench and GPQA-diamond. The card does not state whether thinking mode was enabled for the reported scores. A benchmark chart image accompanies the table.
Safety and usage
The card reports no safety, refusal or dangerous-capability evaluations. A usage-considerations section states that the model "has not been specifically designed or comprehensively evaluated for every possible downstream application" and places responsibility for accuracy, safety, fairness and legal compliance on developers and downstream users.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| FORTE | LongCat-2.0 | 73.2% | thinking mode not stated; measured in-house by Meituan under a unified harness; developer-reported | Evaluation Results |
| FORTE | Gemini 3.1 Pro | 70.3% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| FORTE | GPT-5.5 | 77.8% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| FORTE | Claude Opus 4.6 | 73.2% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| FORTE | Claude Opus 4.7 | 77.6% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| FORTE | Claude Opus 4.8 | 77.2% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| RWSearch | LongCat-2.0 | 78.8% | thinking mode not stated; measured in-house by Meituan under a unified harness; developer-reported | Evaluation Results |
| RWSearch | Gemini 3.1 Pro | 76.3% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| RWSearch | GPT-5.5 | 85.3% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| RWSearch | Claude Opus 4.6 | 81.3% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| RWSearch | Claude Opus 4.7 | 79.3% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| RWSearch | Claude Opus 4.8 | 77.3% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| IFEval | LongCat-2.0 | 90.0% | thinking mode not stated; measured in-house by Meituan under a unified harness; developer-reported | Evaluation Results |
| IFEval | Gemini 3.1 Pro | 96.1% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| IFEval | GPT-5.5 | 95.0% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| IFEval | Claude Opus 4.6 | 92.2% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| IFEval | Claude Opus 4.7 | 88.7% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| IFEval | Claude Opus 4.8 | 86.0% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| Writing Bench | LongCat-2.0 | 83.8% | thinking mode not stated; measured in-house by Meituan under a unified harness; developer-reported | Evaluation Results |
| Writing Bench | Gemini 3.1 Pro | 83.7% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| Writing Bench | GPT-5.5 | 84.7% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| Writing Bench | Claude Opus 4.7 | 85.3% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| Writing Bench | Claude Opus 4.8 | 85.2% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| IMO-AnswerBench | LongCat-2.0 | 81.8% | thinking mode not stated; measured in-house by Meituan under a unified harness; developer-reported | Evaluation Results |
| IMO-AnswerBench | Gemini 3.1 Pro | 90.0% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| IMO-AnswerBench | GPT-5.5 | 79.5% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| IMO-AnswerBench | Claude Opus 4.6 | 75.3% | thinking mode not stated; cited from the model's official report, as reprinted in the LongCat-2.0 model card | Evaluation Results |
| IMO-AnswerBench | Claude Opus 4.7 | 81.8% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
| IMO-AnswerBench | Claude Opus 4.8 | 75.3% | thinking mode not stated; measured in-house by Meituan under its unified harness; reported by Meituan in the LongCat-2.0 model card | Evaluation Results |
Provenance
The page summarises the Hugging Face card as captured on September 21, 2026. All figures are reported by Meituan; comparison figures are either Meituan's own runs or the other developers' published numbers, as marked. Confidence is medium: a single developer-authored source with fast-decay benchmark claims and no third-party evaluation.
Relationships
- supports: LongCat-2.0 (Meituan)
- instance-of: AI System Cards, Open-Weight Frontier Models
- related: Meituan / LongCat Team, Terminal-Bench, SWE-Bench Pro, SWE-bench, BrowseComp, GPQA, Gemini 3 / Gemini 3 Pro, GPT-5.5 ('Spud'), Claude Opus 4.8