Meituan's LongCat-2.0 breaks the ceiling for open-source coding models, featuring a 1.6T MoE architecture and outperforming GPT-5.5 on SWE-bench Pro.
On June 30, 2026, the landscape of artificial intelligence shifted. Meituan has officially released LongCat-2.0, a milestone coding model that marks a historical turning point in the democratization of high-end AI. For months, the developer community has been testing the 'Owl Alpha' model via OpenRouter, but the full, unbridled power of the 1.6T parameter powerhouse is now available to the open-source community.
This isn't just another incremental update. LongCat-2.0 represents a leap in how massive-scale Mixture-of-Experts (MoE) models can be optimized for real-world software engineering. By bridging the gap between proprietary giants and open-source accessibility, Meituan has set a new benchmark for what developers can expect from local and hosted autonomous agents.
At the heart of LongCat-2.0 lies a sophisticated Mixture-of-Experts architecture. While the total parameter count sits at a staggering 1.6 trillion, the model is designed for extreme efficiency. Through its specialized design, it maintains approximately 48B active parameters per token, ensuring that the massive scale does not result in prohibitive latency.
The model introduces the MOPD (Multi-Objective Parameter Distribution) architecture, which utilizes three specialized expert groups: Agent, Reasoning, and Interaction. These groups are gate-routed dynamically based on the specific task at hand. This is further enhanced by 'Zero-Compute Experts,' a mechanism that allows the model to dynamically activate between 33B and 56B parameters per token, ensuring that no compute cycle is wasted on irrelevant expert pathways.
To handle massive codebases, Meituan implemented LongCat Sparse Attention (LSA). This innovation allows the model to scale efficiently across a massive 1-million-token context window, making it possible to ingest entire repositories, extensive documentation, and complex dependency trees in a single inference session.
The performance metrics for LongCat-2.0 are nothing short of industry-shaking. In the realm of software engineering, the model achieved a score of 59.5 on SWE-bench Pro, officially surpassing GPT-5.5's score of 58.6. This demonstrates a superior ability to handle complex, real-world repository-level issues autonomously.
LongCat-2.0 also shows incredible versatility in multilingual environments and specialized technical tasks. It scored 77.3 on SWE-bench Multilingual, proving its utility for global development teams. Furthermore, its performance on specialized reasoning and browsing benchmarks shows a well-rounded intelligence that extends beyond simple code completion.
Technical performance summary:
Terminal-Bench 2.1: 70.8
FORTE: 73.2
RWSearch: 78.8
BrowseComp: 79.9
Meituan has optimized the pricing structure to encourage large-scale deployment, particularly for RAG (Retrieval-Augmented Generation) and long-context agentic workflows. The inclusion of aggressive cache-hit pricing makes it highly economical for developers working with repetitive large-scale contexts.
With a 1-million-token context window, the cost-to-performance ratio is designed to compete directly with top-tier proprietary APIs while offering the flexibility of an open-source backbone.
LongCat-2.0 is engineered for high-complexity environments. Its primary strength lies in Autonomous Software Engineering Agents. Because it can hold a million tokens in its 'working memory,' it can act as a true junior engineer—reading documentation, navigating files, and executing terminal commands without losing the thread of the task.
Beyond coding, the MOPD architecture makes it an ideal candidate for advanced Reasoning and Interaction tasks. Whether you are building a complex RAG system that requires deep semantic understanding of massive datasets or an interactive agent that must reason through multi-step logical problems, LongCat-2.0 provides the necessary depth and scale.
Developers can begin integrating LongCat-2.0 immediately. The model weights are being released through official Meituan channels for local deployment, and the API is available via standard endpoints for those preferring managed infrastructure.
For those looking to experiment with the scale of the model before full deployment, the OpenRouter integration provides a seamless way to test the capabilities of the underlying architecture.
API Pricing — Input: 0.75 / Output: 2.95 / Context: 1000000