Meituan Open Sources LongCat-2.0, Simultaneously Releases Inference Code for Domestic Chips

Meituan LongCat-2.0 Officially Released: Trillion-Parameter Model Completes Full-Process Training and Inference on Domestic Compute Cluster

On June 30, Meituan officially released the new generation trillion-parameter large model LongCat-2.0 and made it open source.

As the industry's first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic compute cluster (total parameters 1.6T, average active approximately 48B, dynamic range 33B~56B), LongCat-2.0 was pre-trained from scratch and natively supports 1M ultra-long context. Its architecture design revolves around one core goal: enabling the model to complete code understanding, generation, and execution more efficiently and stably in real-world Agentic Coding tasks.

Before the official release, the LongCat-2.0 preview version was opened to global developers via the OpenRouter platform and longcat.ai. As of now, the model has ranked in the top three for global large model call volume on OpenRouter, with monthly call volume ranking first, second, and third globally in Hermes, Claude Code, and OpenClaw respectively, becoming one of the most popular models among global Agent developers.

01 Domestic Model and Chip Full-Stack Collaboration: Completing Stable Training of Trillion-Parameter MoE Model on Domestic Compute

The LongCat team's exploration of domestic compute began in 2023. Over three years, the team started from a thousand-card scale, gradually overcoming basic challenges such as operator adaptation, communication optimization, and distributed stability, ultimately completing full-process training and inference of a trillion-parameter model on a 50,000-card cluster.

LongCat-2.0 pre-training data scale exceeds 30T tokens, covering Chinese, English, multilingual, and code data. Facing hardware failures, communication anomalies, memory pressure, and numerical fluctuations in ten-thousand-card level training, the LongCat team tackled domestic compute training challenges from three aspects: stability, correctness, and efficiency.

- In stability, by handling inter-card communication anomalies, elastic scaling, and automatic fault recovery, the average daily fault rate was reduced by over 70%.

- In correctness, by designing deterministic operators, Bitwise consistency verification, and parameter detection, the reliability of training results was ensured, while improving calculation precision of key modules and optimizing Reduce logic based on practice.

- In efficiency, through pipeline scheduling, memory optimization, and operator-level core control, training MFU was improved by 1.5 times.

Ultimately, LongCat achieved a stable daily throughput exceeding 1T tokens/day, completing stable training of a trillion-parameter MoE model on domestic compute.

During the inference phase, LongCat-2.0 performed collaborative optimization around the model, operators, and framework: supporting low-latency decoding for trillion-parameter MoE models by aggregating memory bandwidth through large-scale expert parallelism; integrating the zero-compute expert mechanism into the expert parallel communication process to truly avoid unnecessary transmission and computation for tokens routed to zero experts; and optimizing scheduling for core operators such as communication, Attention, and GEMM, combined with framework mechanisms like early dispatch and weight prefetching, further reducing waiting overhead in the inference link.

From stable training to low-latency inference, LongCat-2.0 verifies that we possess the capability to train large-scale models on domestic compute clusters. It is not just about "being able to train" a trillion-parameter model, but also enabling trillion-parameter models to run stably in real tasks.

02 Enabling the Model to Complete Code Understanding, Generation, and Execution More Efficiently and Stably in Real Agentic Coding Tasks

LongCat-2.0's architecture design always revolves around one core goal: enabling the model to complete code understanding, generation, and execution more efficiently and stably in real Agentic Coding tasks.

1M Ultra-Long Context, Letting Agents See the Whole Project. Traditional models start "forgetting" previous content after processing over 100K context. LongCat-2.0 adopts the LongCat Sparse Attention (LSA) sparse attention mechanism. When processing long text, it no longer "reads word by word," but intelligently filters key information, reducing computation from square level to linear level. This allows the model to maintain precise information positioning and understanding capabilities even within a 1 million token ultra-long context.

Zero Compute Experts + ScMoE, Putting Compute Where It Counts. In coding tasks, the complexity of different tokens varies greatly—the compute demand for defining variable names differs completely from deriving recursive algorithms. LongCat-2.0 achieves token-level dynamic activation (33B~56B) through zero compute experts; simple tokens do not consume compute, while complex tokens automatically receive more computational resources.

MOPD Multi-Expert Fusion, One Model Excelling at Coding, Reasoning, and Interaction. LongCat-2.0 fuses three sets of expert capabilities—Agent, Reasoning, and Interaction—through the MOPD architecture. Agent Experts specialize in tool calling and autonomous error correction; Reasoning Experts delve into math and STEM reasoning; Interaction Experts optimize instruction following and interaction experience. During inference, a gating network dynamically schedules the most suitable experts based on task type, rather than simply merging parameters. Thanks to this, the model performs outstandingly across dimensions such as programming, reasoning, and interaction.

Through refined architecture design, LongCat-2.0 enables trillion-parameter models to exert their capabilities more efficiently and stably in actual tasks.

03 Excellent Performance in Programming Capabilities and Complex Task Processing in Real Office Scenarios

Comprehensive evaluation results show that LongCat-2.0 performs excellently in Code and General Agent scenarios thanks to its outstanding comprehensive performance and stable task execution.

- In programming capabilities, LongCat-2.0 demonstrates solid comprehensive strength: scoring 59.5 in SWE-bench Pro, which tests deep engineering capabilities, leading Gemini 3.1 Pro (54.2), GPT-5.5 (58.6), and Claude Opus 4.6 (57.3); achieving 77.3 in SWE-bench Multilingual, keeping pace with Claude Opus 4.6 (77.8); additionally, scoring 70.8 in the real terminal instruction interaction evaluation Terminal-Bench 2.1, reflecting its stable execution and error correction capabilities in real operations and development terminal tasks.

- In complex task processing in real office scenarios, LongCat-2.0 performs balanced: scoring 78.8 in the search agent evaluation set RWSearch, 73.2 in the productivity scenario evaluation set FORTE, and 79.9 in BrowseComp, all reaching or approaching the level of frontier closed-source models. This proves its high reliability in multi-step task planning, complex tool calling, and long-range retrieval execution, fitting the landing needs of enterprise-level Agents well.

04 Becoming a Reliable "Work Partner" in Real Work Scenarios

During the internal testing period, we collected a large number of real user task requirements facing real work scenarios. These real "work orders" from the front line show that LongCat-2.0 is becoming a reliable "work partner" in users' real work scenarios.

Agent Building: One Question, Full Closed-Loop Delivery

Through an AI SQL Agent built with LongCat-2.0, business personnel can directly query data using natural language. LongCat-2.0 automatically completes the full link closed-loop—understanding problem intent, planning query steps, and converting data results into clear business insights.

Codebase Migration: Understanding Old Code, Refactoring New Architecture

Given an old plugin codebase and a new SDK document, LongCat-2.0 can independently analyze the overall architecture, sort out core logic, and refactor the entire plugin into an implementation conforming to the new API—retaining all original functions, fixing potential hazards, and passing compilation on the first try.

Complete Application Development: From One Sentence to a Runnable Product

Describing a "Children's AI Game Training Ground" concept, LongCat-2.0 will gradually generate technology selection, page architecture, game logic, and visual details—from the homepage to three complete playable game pages, all code produced at once, ready to use. From one sentence to a usable product, easily realizing inspiration.

3D Interaction Demo: One Sentence, Generate a 3D World

Through a one-sentence description, LongCat-2.0 can generate a complete Three.js 3D demo: transparent flask, fluorescent liquid, foam eruption, liquid level drop, and accumulation effects are all interactively presented. All code is encapsulated in one HTML file, open and use, allowing creativity to quickly transform into an interactive 3D experience.

AI Novel Factory: From Single Point Inspiration to Commercial Monetization

Based on LongCat-2.0, the "AI Novel Factory" upgrades creative writing into an automated content pipeline. After users input inspiration, the system orchestrates multiple Agents to automatically complete world-building, parallel chapter generation, quality assessment, and feedback revision. Long-context capabilities ensure consistency for million-word level settings. Final content can automatically adapt to multi-platform publishing, and a Web panel monitors generation progress and quality status in real-time, achieving continuous and stable serialized output.

Try Now.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.