LongCat-2.0 Official Model Card: Specifications and Official Benchmarks (Including Comparison Tables with Gemini/GPT-5.5/Claude Opus)

One-sentence takeaway

The official model card is the primary authoritative source for judging LongCat-2.0's suitable tasks: it scores 59.5 on SWE-bench Pro, ahead of GPT-5.5 (58.6) and Gemini 3.1 Pro (54.2), and reaches 70.8 on Terminal-Bench 2.1. However, it trails GPT-5.5 and Claude Opus 4.8 on several benchmarks including BrowseComp, GPQA, and IFEval—in short, it is strong at coding and agent tasks, but not a leader in retrieval and general reasoning.

Official benchmark data (model card's original table; * cites the other model's official report, while unmarked scores were self-tested by Meituan using a unified harness)

Note: * = cited from the corresponding model's official report; - = no publicly comparable score.

Official specification highlights

• Architecture: MoE, 1.6T total parameters, about 48B active per token (dynamic 33B–56B, according to the official X account); HF metadata lists a 1.8T model size (including 135B N-gram Embedding parameters, BF16).

• Context: Native 1M tokens (trained on hundreds of billions of tokens of million-token-context data).

• Key features: LongCat Sparse Attention (LSA, with SI/CLI/HI optimizations), 3-step MTP speculative decoding, and N-gram Embedding (n-gram size=5, 135B parameters).

• Training: 35T+ tokens and more than 50,000 domestic compute chips (AI ASIC superpods), with no rollback throughout (according to the official blog).

• Supported harnesses: Claude Code, OpenClaw, Hermes (officially described as “deeply integrated”).

• License: MIT.

• Official experience entry point: https://longcat.ai; contact longcat-team@meituan.com.

Verification and applicability boundaries

• Methodology warning: All scores are Meituan self-tests (a self-built harness and in-house scoring; the model card notes “problematic tasks corrected”—AlphaSignal points out that this means problematic tasks were manually corrected before scoring). Scores marked * come from each provider's official report, so they are not directly comparable across laboratories; as of 2026-07-01, no independent third party (Artificial Analysis, Scale, etc.) had published a retest score for LongCat-2.0 (according to AlphaSignal's original text).

• Relative strengths and weaknesses (consistent across multiple independent interpretations):

  • Areas ahead: SWE-bench Pro 59.5 > GPT-5.5 58.6 and > Gemini 3.1 Pro 54.2; Terminal-Bench 2.1 is tied with Gemini 3.1 Pro (70.8 vs. 70.7); RWSearch and FORTE exceed Gemini 3.1 Pro.

  • Areas behind: FORTE/RWSearch/BrowseComp all trail GPT-5.5 (77.8/85.3/84.4 vs. 73.2/78.8/79.9); GPQA-diamond and IFEval trail GPT-5.5 and Gemini 3.1 Pro; Claude Opus 4.8 leads LongCat by nearly 10 points on SWE-bench Pro (69.2).

  • Overall: In the official chart, the only model LongCat-2.0 consistently beats is Gemini 3.1 Pro (AlphaSignal's conclusion); against GPT-5.5 it narrowly wins only on SWE-bench Pro.

• Task-fit judgment: Suitable for repository-level coding, terminal interaction, and long-context agent workflows; unsuitable for scenarios requiring top-tier retrieval (BrowseComp), scientific reasoning (GPQA), and strict instruction following (IFEval).

• Related documents: Review 02 (official technical blog) and Review 03 (OpenRouter third-party methodology).