Meituan Officially Launches the Trillion-Parameter Open-Source Large Model LongCat-2.0 with Native Support for 1M Ultra-Long Context

On June 30, Meituan officially released and open-sourced its new trillion-parameter large model, LongCat-2.0. As the first in the industry to complete full-cycle training and inference on a domestic computing cluster with 50,000 cards, LongCat-2.0 has 1.6T total parameters (average activation of about 48B, dynamic range from 33B to 56B), and natively supports 1M ultra-long context.

After its preview version was released on the OpenRouter platform, its monthly usage has ranked among the top three globally, performing outstandingly in ecosystems such as Hermes and Claude Code, becoming one of the most popular agent models for developers worldwide.

Its release marks a significant breakthrough in domestic computing power for large-scale cluster training. Since 2023, the LongCat team has spent three years overcoming fundamental challenges such as operator adaptation, communication optimization, and distributed stability. Through self-developed deterministic operators and elastic recovery mechanisms, they have reduced the average daily failure rate by more than 70%, achieving a stable daily throughput of over 1T tokens.

In terms of architecture design, LongCat-2.0 is centered around real Agentic Coding tasks. It introduces a sparse attention mechanism (LSA) to reduce long text computation to a linear level, and uses a zero-computation expert mechanism combined with MOPD multi-expert fusion architecture to achieve token-level dynamic activation. This enables the model to perform exceptionally well in complex office scenarios such as code understanding, mathematical reasoning, and long-range retrieval. In authoritative programming evaluation benchmarks like SWE-bench Pro, it even surpasses GPT-5.5 and Claude Opus4.6, further accelerating the closed-loop implementation and application restructuring of enterprise-level AI Agents.