LongCat-2.0

A model documented in the meituan-longcat GitHub organization and exposed through the LongCat publisher API platform.

Independent third-party resource. Not affiliated with or endorsed by LongCat or Meituan.

Overview

LongCat-2.0 is a 1.6-trillion-parameter Mixture-of-Experts (MoE) language model built by Meituan's LongCat team. It is the industry's first trillion-parameter model to complete both full training and inference without NVIDIA GPUs — running instead on a cluster of 50,000 domestic AI ASICs. The model activates approximately 48 billion parameters per token (dynamic range 33B–56B) and natively supports a 1-million-token context window through LongCat Sparse Attention (LSA). It was released in June 2026 after operating anonymously on OpenRouter as 'Owl Alpha' for approximately two months, during which it ranked among the global top three by daily token volume and ranked first on Hermes Agent, second on Claude Code, and third on OpenClaw by monthly call volume. LongCat-2.0 is available under the MIT License, with weights on HuggingFace and access through the LongCat API (OpenAI and Anthropic compatible) and OpenRouter.

Architecture

LongCat-2.0 introduces three major architectural innovations. First, the Shortcut-connected Mixture of Experts (ScMoE) design enlarges the computation-communication overlap window, enabling training at tens-of-thousands-of-accelerators scale while maintaining inference throughput. Second, Zero-Computation Experts introduce a dummy expert that handles simple tokens with zero compute cost — a PID controller dynamically adjusts expert bias to maintain an average activation of approximately 48 billion parameters per token. Third, the N-gram Embedding module adds 135 billion additional parameters through 5-gram token combinations, improving parameter utilization without increasing per-token compute. For attention, LongCat Sparse Attention (LSA) uses three-tier indexing — Streaming-aware Indexing, Cross-Layer Indexing, and Hierarchical Indexing — to reduce attention complexity from quadratic to near-linear, enabling the 1-million-token native context. The model also employs MOPD (Multi-Teacher On-Policy Distillation), which fuses Agent, Reasoning, and Interaction expert clusters trained under separate teacher policies into a single model, described by the publisher as enabling strong performance across agentic coding, tool use, and interactive dialogue without task-specific routing.

Training

LongCat-2.0 was pre-trained on over 35 trillion tokens. The publisher reports that this is the first trillion-parameter model to complete full training and inference end-to-end on domestic AI hardware — a cluster of approximately 50,000 AI ASICs — without any NVIDIA GPUs. The ScMoE architecture with Zero-Computation Experts was designed to make this scale of training feasible on non-NVIDIA hardware. Pre-training was followed by the DORA (Dynamic ORchestration for Asynchronous rollout) reinforcement learning framework, which supports large-scale multi-environment training across over 10,000 environments spanning more than 20 domains, with a reported threefold training speedup over synchronous methods.

Key Benchmarks

Publisher-reported scores. Not independently verified by this site. Descriptions summarize what each benchmark measures.

BenchmarkScoreWhat It Measures
SWE-bench ProSoftware engineering tasks on real GitHub repositories — measures ability to understand and modify complex codebases
SWE-bench MultilingualMulti-language variant of SWE-bench, covering non-English programming tasks
Terminal-Bench 2.1Real terminal command execution with error recovery — measures system-level agentic capability
BrowseCompComplex web browsing and multi-page information retrieval under time constraints
RWSearchReal-world search agent evaluation — measures ability to find, filter, and synthesize web information
FORTEProductivity and office scenario tasks — measures performance on real-world business workflows

Deployment

LongCat-2.0 model weights are distributed on HuggingFace (meituan-longcat organization) in BF16, FP8, and INT8 precision variants. The LongCat API platform at longcat.chat provides both OpenAI-compatible (api.longcat.chat/openai/v1/chat/completions) and Anthropic-compatible (api.longcat.chat/anthropic/v1/messages) endpoints. On OpenRouter, the model is listed under Meituan LongCat 2.0 with pricing at $0.30/M input and $1.20/M output tokens (as of July 2026). For self-hosted deployment, SGLang PR #30042 adds dedicated GPU inference support with a recommended configuration of 16x H20 GPUs using tensor and expert parallelism. A separate NPU inference repository (meituan-longcat/SGLang-FluentLLM) provides optimized inference on domestic AI accelerators. The model is also compatible with vLLM and Hugging Face Transformers (trust_remote_code=True).

Capabilities

Described in the cited primary documentation. Not an independent performance evaluation.

  • Documented in the meituan-longcat GitHub repository.
  • Listed in the publisher API documentation as a supported model.
  • Exposed through the publisher API platform in both OpenAI-compatible and Anthropic-compatible formats.

Access & License

Open source. Review the LICENSE file in the meituan-longcat GitHub repository for the full license terms before any commercial use.

Documented context window: 1,000,000 tokens (per the cited publisher API documentation).

Sources

  • LongCat-2.0 GitHub repository

    Primary sourceAccessed 2026-07-17

    Primary repository under the meituan-longcat organization; contains the model code, configuration, and license file.

  • LongCat API Quick Start

    Publisher documentationAccessed 2026-07-17

    Publisher documentation listing the supported models and the chat completion endpoints.

  • LongCat API overview

    Publisher documentationAccessed 2026-07-17

    Publisher documentation describing the base URL, authentication, endpoints, error responses, and SDK compatibility.

Related guides

  • LongCat API Quick Start

    A primary-source-based introduction to creating a LongCat API key and sending a basic chat completion request through the OpenAI-compatible endpoint.

Related comparisons

FAQ

Where can the primary source for LongCat-2.0 be found?
The official GitHub repository (meituan-longcat) and HuggingFace page contain the primary documentation, model weights, and license for LongCat-2.0.
Is LongCat-2.0 available through the LongCat API?
Yes. LongCat-2.0 is listed in the publisher API documentation on both OpenAI-compatible and Anthropic-compatible endpoints.
Is this page an official LongCat or Meituan page?
No. This page is published by an independent third-party site. It is not affiliated with, endorsed by, or sponsored by LongCat or Meituan.
How often is this page updated?
This page was last verified on 2026-07-17. Content is reviewed when new publisher documentation or model releases become available.

Independent third-party disclosure

This page is published by an independent third-party site. It is not affiliated with, endorsed by, sponsored by, or operated by Meituan, LongCat, or any of their affiliates. The content summarizes publicly-available primary documentation and does not represent the views of any referenced organization.

Last reviewed: 2026-07-17