Wait Which Model?
← Back to directory
Meituan (LongCat)

LongCat-2.0

Released Jun 29, 2026 · knowledge cutoff unpublished

Status
Superseded
Location
China
Type
LLM
Modality
Text
Context window
1M
Max output
262K
Speed
42 tok/s · 2.78s to first answer token
Price ($/MTok in / out)
$0.75 / $2.95
Cost per task
$0.12
Open weights
Yes
Benchmarks4 of 10 reported

GPQA Diamond

88.9%

SWE-bench Pro

59.5%

Terminal-Bench 2.1

70.8%

Humanity's Last Exam

32.1%

MMLU-Proretired

—

SWE-bench Verified

—

AIME (latest)retired

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Trained and served end to end on domestic Chinese accelerators, with no Nvidia silicon anywhere in the loop
  • Spent two months at the top of OpenRouter as the anonymous 'Owl Alpha' before anyone knew whose it was — the reception came before the branding
  • Careful and thorough on review-style work where you want it to explain what it found
  • MIT-licensed at trillion-parameter scale, with a clean commercial grant

Weaknesses

  • Verbose and slow enough to be awkward for conversational pairing — it suits long unattended runs better
  • Hands-on comparisons put its generated apps a generation behind GLM 5.2's: rougher and buggier despite benchmark parity
  • Fabricates readily on factual questions, and knowledge coverage lags its agentic ability badly

Announced 2026-06-29 on GitHub, Hugging Face and Meituan's own platform (some coverage dates it 2026-06-30); it had been running anonymously on OpenRouter as 'Owl Alpha' for roughly two months beforehand. GPQA Diamond 88.9 and Terminal-Bench 2.1 70.8 are Meituan's in-house harness figures — Artificial Analysis independently measures 78.0 and 50.2 on the same two benchmarks, and HLE 32.1% is Artificial Analysis' no-tools run since Meituan published none. Also reported: SWE-bench Pro 59.5%, SWE-bench Multilingual 77.3%, FORTE 73.2%, BrowseComp 79.9%; no SWE-bench Verified, MMLU-Pro, AIME or ARC-AGI-2 figure exists. Pricing is LongCat's standard API rate, with a launch promotion at $0.30/$1.20. Max output is the OpenRouter endpoint spec; some coverage cites 131,072. Meituan says only 'AI ASIC superpods' — the widely repeated 50,000 Huawei Ascend 910C figure comes from Chinese media, not from Meituan. Speed is Artificial Analysis' median across providers serving LongCat 2.0.

For developers

API model strings

  • longcatLongCat-2.0

Licence

MIT License · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor

News