AB
AB
AiBoss

All Categories

project

LongCat-2.0 - Meituan's open-source next-generation trillion-parameter language model

LongCat-2.0 is a large-scale MoE language model open-sourced by Meituan, with a total of 1.6 trillion parameters. Each token activates approximately 48 billion parameters, and training and deployment are completed based on an AI ASIC supercomputing cluster.

What is LongCat-2.0?

LongCat-2.0 is a large-scale MoE language model open-sourced by Meituan, boasting 1.6 trillion total parameters, with approximately 48 billion parameters activated per token. It was trained and deployed using an AI ASIC supercomputing cluster. The model supports ultra-long contexts of up to 1MB and incorporates LongCat Sparse Attention and N-gram Embedding architectures. It delivers strong performance in code generation, agent task execution, and complex inference scenarios, and is deeply integrated with mainstream tools such as Claude Code, OpenClaw, and Hermes.

Main functions of LongCat-2.0

  • Extremely long context understandingIt natively supports a 1M token context window, which can read the entire codebase and long documents at once for accurate positioning and in-depth understanding, avoiding the "forgotten in the middle" problem of traditional models.
  • Code generation and refactoringBased on Agentic Coding capabilities, it supports repository-level code analysis, cross-file logic sorting, architecture migration and function refactoring, and compiles successfully on the first try.
  • Autonomous Agent ExecutionIt deeply integrates with mainstream harnesses such as Claude Code, OpenClaw, and Hermes, and can autonomously complete multi-step task planning, tool calls, API interaction, and error self-correction.
  • Multilingual code supportIt has the ability to understand and generate cross-language code, and supports code migration and review in programming scenarios involving multiple languages such as Chinese and English.
  • Intelligent Search and RetrievalBuilt-in search and browsing capabilities support complex information retrieval, multi-hop reasoning, and cross-web page content integration, meeting the needs of research tasks.
  • Natural Language Data QueryBusiness personnel can directly query the database using natural language, and the model automatically completes intent parsing, SQL generation, execution, and result insight output.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

Technical Principles of LongCat-2.0

  • MoE Sparse ArchitectureIt adopts a hybrid expert architecture with a total of 1.6 trillion parameters and approximately 48 billion activations per token, achieving a balance between parameter scale and computational efficiency through sparse activation.
  • LongCat Sparse Attention (LSA)The introduction of three major components—Streaming Aware Index (SI), Cross-Layer Index (CLI), and Hierarchical Index (HI)—reduces the computation of long context attention from quadratic to linear, thus solving the bottleneck of processing millions of token contexts.
  • N-gram EmbeddingConfigure the 5-gram embedding module to expand the embedding space by about 100 times through token combination, capture richer local context and improve representation capabilities.
  • Zero Computing Expert + ScMoEComputational resources are dynamically allocated based on token complexity. Simple tokens are routed to zero experts to avoid ineffective computation, while complex tokens automatically receive more expert resources.
  • MOPD multi-expert fusion trainingThe system integrates three types of expert groups—Agent Experts, Reasoning Experts, and Interaction Experts—through a Multi-Teacher On-Policy Distill architecture, with the gating network dynamically scheduling the optimal expert combination based on the task type.
  • 6D Parallel TrainingBased on the standard TP/CP/EP/DP/PP, EMBP is added, and combined with the Superpod physical cluster architecture, it enables ultra-large-scale and efficient training.

How to use LongCat-2.0

  • Online experienceVisit the LongCat official website to directly interact with and try out the model capabilities.
  • API AccessLog in to https://longcat.chat/platform/product to obtain the API Key and integrate it into your own application or Agent framework.
  • Tool IntegrationIt directly integrates with mainstream AI development toolchains such as Claude Code, OpenClaw, and Hermes as the underlying model driver.

LongCat-2.0's core advantages

  • The entire chain of domestic computing powerThe industry's first model to complete the entire training and inference process of a trillion-parameter model on a domestic cluster with 50,000 cards, verifying the feasibility of non-NVIDIA platforms.
  • Millions of contexts:1M native context window, combined with LSA sparse attention, enables accurate positioning and understanding of long texts, eliminating "forgetting in the middle".
  • Agentic Coding DedicatedWith in-depth optimizations in code understanding, generation, and execution, it surpasses Gemini 3.1 Pro and GPT-5.5 in benchmarks such as SWE-bench and Terminal-Bench.
  • Dynamic calculation allocationZero-computation expert + ScMoE enables token-level dynamic activation (33B~56B), simple tokens consume no computing power, and complex tasks automatically obtain more resources.
  • Three experts integrationThe MOPD architecture integrates Agent, Reasoning, and Interaction expert groups, and uses a gated network for dynamic scheduling, taking into account execution, reasoning, and interactive experience.
  • Global Developer RecognitionThe preview version is available for use through OpenRouter, and its monthly usage ranks among the top three globally, ranking first, second, and third among Hermes, Claude Code, and OpenClaw, respectively.

LongCat-2.0 project address

  • Project official website: https://longcat.chat/blog/longcat-2.0/
  • GitHub repository: https://github.com/meituan-longcat/LongCat-2.0
  • HuggingFace model libraryhttps://huggingface.co/meituan-longcat/LongCat-2.0

Comparison of LongCat-2.0 with similar competing products

Comparison Dimensions LongCat-2.0 DeepSeek-V3
Publisher Meituan DeepSeek
Total parameters 1.6 trillion (MoE) 671 billion (MoE)
Activation parameters ~48 billion/token ~37 billion/token
Context length 1M (1 million) tokens 64K-128K tokens
Training hardware 50,000-card domestic computing power cluster NVIDIA GPU cluster
Attention mechanism LongCat Sparse Attention (LSA) Multi-head Latent Attention (MLA)
Embedding optimization N-gram Embedding (5-gram, 135B) Standard Embedded Layer
Post-training architecture MOPD (Multi-Expert Fusion) (Agent/Reasoning/Interaction) Standard SFT + RL
Coding capabilities (SWE-bench Pro) 59.5 ~50+

Application scenarios of LongCat-2.0

  • Large-scale code migrationIt reads the complete codebase and migration documentation, automatically maps the architecture and refactors plugins to the new SDK, preserving functionality and fixing potential bugs.
  • Repository-level code reviewBased on a 1M context, we can understand the entire project structure and perform cross-file bug detection, performance optimization, and architecture suggestions.
  • AI Agent DevelopmentAs the underlying model that drives frameworks such as Claude Code and OpenClaw, it executes complex multi-step tasks and autonomous tool calls.
  • Long document analysis and generationProcessing technical documents, papers, and reports of millions of words, performing in-depth summarization, Q&A, and content rewriting.
  • Data Query and InsightBusiness users query the database using natural language, and the model automatically generates, executes, and interprets the SQL queries.
Back