Skip to main content

LongCat

Category: Foundation Models / LLMs

Meituan's open-source large language model family, featuring the first trillion-parameter model (LongCat-2.0) trained and deployed entirely on domestic Chinese AI chips. LongCat was founded in 2023. The company is led by Wang Xing (Meituan CEO); Pei Peng (LongCat basic model lead); Pan Xin (multimodal AI lead). Based in Beijing, China. Team size: 101-500. Latest round: Undisclosed. Key investors include Meituan (parent company, publicly traded HKEx:3690).

Founded
2023
Headquarters
Beijing, China
Team size
101-500

Value proposition

Open-source, high-performance AI models trained entirely on domestic Chinese AI chips (zero NVIDIA dependency), offering competitive performance against frontier models at significantly lower cost ($0.75/$2.95 per million tokens vs GPT-5.5's $5/$30).

Products and solutions

LongCat-2.0 (1.6T MoE, agentic coding model, 1M context, open-source MIT), LongCat-Flash-Chat (560B MoE, 27B activated, 128K context), LongCat-Next (native discrete multimodal model unifying image/audio/text), LongCat-Flash-Thinking-2601 (reasoning model with Re-thinking Mode), LongCat-Video (13.6B DiT video generation), LongCat-Image (6B image generation), LongCat-Flash-Prover (theorem proving), LongCat-Video-Avatar 1.5 (digital human video), LongCat-AudioDiT (audio generation)

Unique value

First trillion-parameter model achieving full training-to-inference闭环 on domestic Chinese ASICs (50,000+ card cluster), with zero NVIDIA dependency, while delivering near-frontier performance at a fraction of the cost.

Target customer

Global AI developers, enterprises needing cost-effective coding/agentic AI, Chinese enterprises seeking sovereign AI infrastructure, open-source community

Industries served

Software Development, E-commerce/Delivery (Meituan internal), Enterprise AI, Agentic AI, Multimodal AI applications

Technology advantage

LongCat Sparse Attention (LSA) for linear-complexity 1M context; Zero-Computation Experts for token-level dynamic compute; ScMoE (Sparse Communication MoE) for reduced cross-node communication; N-gram Embedding for improved token-level pattern recognition; MOPD (Multi-Teacher On-Policy Distill) for multi-expert fusion; DORA (Dynamic ORchestration for Asynchronous rollout) RL training system; full domestic chip training pipeline with automated fault handling (reduced daily fault rate from 15.7% to 4.4%)

How they differentiate

Unlike DeepSeek (which used NVIDIA for pre-training and domestic chips only for inference), LongCat-2.0 is the first trillion-parameter model to complete both pre-training and inference entirely on domestic Chinese ASICs. It also differentiates through aggressive pricing (3-10x cheaper than comparable frontier models), full open-source MIT licensing, and proven market validation via anonymous deployment (Owl Alpha) that topped OpenRouter rankings before Meituan revealed its identity.

Main competitors

DeepSeek (DeepSeek V4-pro/R1), Alibaba Cloud (Qwen3 family/Qwen2.5), Moonshot AI (Kimi K2)

Key partnerships

OpenRouter (model hosting/distribution), Hugging Face (model distribution), GitHub (open-source code), ModelScope (model distribution), Shanghai Jiao Tong University (research collaboration), Hong Kong University of Science and Technology (research collaboration)

Notable customers

Global developer community on OpenRouter (ranked #1 on Hermes Agent, #2 on Claude Code, #3 on OpenClaw by monthly call volume), Meituan internal teams (95%+ code work uses internal CatPaw tool powered by LongCat)

Major milestones

2023-01: LongCat base team established, began building domestic compute cluster, 2023: Ran first thousand-parameter training on domestic chips, 2024: Validated MoE architecture on domestic chips, 2025-09: Released LongCat-Flash-Chat (560B MoE, first public release), 2025-12: Released LongCat-Flash-Thinking, 2026-01: Released LongCat-Flash-Thinking-2601, 2026-03: Released LongCat-Flash-Omni, 2026-04: Released LongCat-Next (native multimodal), 2026-04: Released LongCat-2.0-Preview, 2026-06-30: Officially released LongCat-2.0 (1.6T, first trillion-parameter model fully trained on domestic chips), 2026-07: Revealed as Owl Alpha, which had been #1 on OpenRouter Hermes Agent for 2 months

Growth metrics

LongCat-2.0: 1.6T total parameters, ~48B activated per token, 1M context window, trained on 35T+ tokens on 50,000+ domestic ASICs. SWE-bench Pro 59.5. Terminal-Bench 70.8. OpenRouter top 3 globally by call volume. HuggingFace: 3,326+ followers, 37+ GitHub repositories.

Market positioning

China's leading sovereign AI model family, positioned as a cost-effective open-source alternative to both US frontier models (GPT-5.5, Claude) and Chinese competitors (DeepSeek, Qwen), with unique differentiation in domestic chip independence and agentic coding performance.

Geographic focus

China (primary), Global (secondary via open-source)

Patents and IP

Open-source under MIT License; multiple technical innovations published in arxiv papers (LongCat Sparse Attention, N-gram Embedding, ScMoE, DORA RL framework, DiNA architecture)

About Wang Xing (Meituan CEO); Pei Peng (LongCat basic model lead); Pan Xin (multimodal AI lead)

Wang Xing: Co-founder and CEO of Meituan (est. 2010), previously founded Xiaonei (Chinese Facebook clone) and Fanfou (microblogging service). Pan Xin: Former Google DeepMind researcher, former head of vision large model at ByteDance, former CTO of Sharge. Pei Peng: Person in charge of Meituan LongCat basic model team.

Official website: