Skip to content
All systems operational0 AI providers monitored, polled every 2 minutes
Live status

Meituan is the Chinese consumer platform (food delivery, travel, local services) whose AI lab published one of the more consequential open-source releases of 2026. On June 30, 2026 Meituan open sourced LongCat-2.0, a 1.6 trillion parameter mixture-of-experts model with dynamic activation of 33 to 56 billion parameters per token, native 1 million token context, and a purpose-built agentic coding orientation. It was the first trillion-parameter release to complete full training and inference entirely on a 50,000-card domestic Chinese compute cluster, a milestone for building frontier AI without leading-edge Western silicon. Weights ship under MIT on GitHub and Hugging Face. Meituan reports 59.5 on SWE-Bench Pro (self-reported), ahead of Gemini 3.1 Pro, GPT-5.5, and Claude Opus 4.6 by their measure; independent verification is pending. A preview version had been quietly running on OpenRouter and longcat.ai for weeks before the announcement, ranking among the top three models globally by call volume during that stealth window. Meituan priced API access to LongCat-2.0 at a list of $0.75 per million input tokens and $2.95 output, discounted for a limited time to $0.30 and $1.20, though the LongCat pay-as-you-go page has since listed LongCat-2.5 alone. On September 25, 2026 Meituan followed with LongCat-2.5-Preview, adding native image input and a 131,072 token output ceiling at the same discounted $0.30 and $1.20 rate, but without the technical report or benchmark table that accompanied LongCat-2.0. The preview has since given way to LongCat-2.5, which is now the only model the LongCat pay-as-you-go API serves: 1.6 trillion total parameters with roughly 48 billion active per token, natively multimodal, 1M context with up to 256K output, OpenAI and Anthropic-shaped endpoints, and MIT weights on Hugging Face alongside a 495B W4A4 quantization, at a limited-time discounted $0.30 per million input tokens, $0.006 cached, and $1.20 output with no standard rate published behind it. Reports put the weights release at November 9 or 10, 2026, roughly a week after the model appeared. Meituan publishes a full benchmark table this time but marks its unmarked figures as in-house and indicative only, with Terminal Bench v2.1 at 88.4 against 89.1 for Claude Opus 5, BrowseComp at 91.9, and Humanity's Last Exam with tools at 63.1; no independent evaluator has tested it. Self-hosting still targets Huawei Ascend hardware rather than NVIDIA.

Founded

2010 (Meituan); 2024 (Meituan AI Lab)

Headquarters

Beijing, China

CEO

Wang Xing

Models

2 active

Key Products

LongCat-2.5LongCat-2.0LongCat platform (longcat.ai)

Strengths

  • ✓1.6T open-source MoE under MIT
  • ✓LongCat-2.5 is the only model on the pay-as-you-go API, at $0.30 / $1.20
  • ✓Trained end-to-end on domestic Chinese silicon
  • ✓Native 1M context
  • ✓Purpose-built for agentic coding
  • ✓Top-three OpenRouter call volume during stealth preview

Meituan Models

ModelInput / 1MOutput / 1MContextCapabilities
LongCat-2.00.752.951Mtext, code, tool-use, reasoning
LongCat-2.50.301.201Mtext, vision, tool-use, code, reasoning

Prices per 1M tokens in USD. See the full pricing guide for detailed analysis.

Benchmark Scores

ModelSWE-benchMMLU-ProHumanEvalGPQA DiamondMATHOSWorld 2.0BrowseCompFrontierCode v1.1Terminal-Bench 4.0Humanity's Last Exam (tools)
LongCat-2.0N/AN/AN/A88.9N/AN/A79.9N/AN/AN/A
LongCat-2.5N/AN/AN/AN/AN/AN/A91.9N/A38.563.1

Other Providers

Visit MeituanCompare ModelsAll Models