Why This Release Matters

LongCat-2.0 is not just another large model release. It is a signal that Chinese AI development can achieve near-frontier capability without access to advanced U.S. GPUs, which are subject to escalating export restrictions . The model demonstrates that a 1.6 trillion-parameter system can be built entirely on domestic silicon, from pre-training through inference.

Technical Specifications

• Architecture: Mixture-of-Experts (MoE) with a total capacity of 1.6 trillion parameters, but only ~48 billion activated per token (sparsity ~97%) .
• Context window: 1 million tokens, enabled by a custom LongCat Sparse Attention (LSA) mechanism that evolved from DeepSeek's sparse attention design .
• Active parameter range: 33B–56B per token via dynamic 'Zero-Compute Expert' routing, ensuring no wasted computation .
• N-gram Embedding: An additional ~135 billion parameters in the embedding layer, using 2-to-5 character n-grams to improve local context representation without stacking more MoE layers .
• Purpose-built for: Agentic coding — code understanding, generation, and execution in real-world tool-use scenarios .
• Training infrastructure: Trained from scratch on 50,000–60,000 Chinese-made AI ASICs (not Nvidia GPUs), consuming over 35 trillion tokens of Chinese, English, multilingual text, and source code .
• License: Released under the MIT open-source license .
• Availability: Weights and code published on GitHub, Hugging Face, and the official LongCat website; also accessible via API and an online demo .

Performance and Benchmarks

Meituan claims LongCat-2.0 achieves performance comparable to Google's Gemini 3.1 Pro . Before its official unveiling, the model operated anonymously as 'Owl Alpha' on OpenRouter, where it reportedly topped developer rankings for coding benchmarks .

Key benchmark scores published by the LongCat team on X include: Terminal-Bench 2.1: 70.8, SWE-bench Pro: 59.5 (with GPT-5.5 at 58.6 for comparison), SWE-bench Multilingual: 77.3, and FORTE: 73.2 .

Strategic Significance

LongCat-2.0 carries implications far beyond benchmark scores:

• First trillion-parameter model trained end-to-end on domestic Chinese chips, from pre-training through inference and deployment — a milestone that even DeepSeek-V4-pro did not achieve, as it relied on domestic hardware only for inference .
• Demonstrates China's ability to achieve near-frontier AI capability without access to Nvidia's advanced GPUs, which are subject to U.S. export restrictions .
• Signals potential decoupling of Chinese AI development from U.S. hardware supply chains, with implications for the global AI chip market and geopolitical tech competition .
• For Meituan specifically, it positions a food-delivery and local-services giant as a serious contributor to open-source AI foundation models, similar to moves by Alibaba (Qwen) and ByteDance .

Key Architectural Innovations

LongCat-2.0 introduces two notable architectural improvements over its predecessor, LongCat-Flash:

1. LongCat Sparse Attention (LSA): An evolution of DeepSeek's sparse attention mechanism (DSA), LSA addresses latency bottlenecks in the indexer through three independent efficiency optimizations: flow-aware indexing, cross-layer indexing, and hierarchical indexing — designed to accelerate long-context processing without sacrificing model quality .

2. MOPD (Multi-Objective Process Decoding): The model organizes its experts into three specialized groups — Agent, Reasoning, and Interaction — with a gate router that directs each token to the appropriate group based on task type .

Availability and Access

Developers and researchers can access LongCat-2.0 under the permissive MIT license. The model weights, inference code, and documentation are available on GitHub, Hugging Face, and the official LongCat website. An API endpoint and an interactive online demo are also provided .

Why This Release Matters

LongCat-2.0 is not just another large model release. It is a signal that Chinese AI development can achieve near-frontier capability without access to advanced U.S. GPUs, which are subject to escalating export restrictions . The model demonstrates that a 1.6 trillion-parameter system can be built entirely on domestic silicon, from pre-training through inference.

Technical Specifications

• Architecture: Mixture-of-Experts (MoE) with a total capacity of 1.6 trillion parameters, but only ~48 billion activated per token (sparsity ~97%) .
• Context window: 1 million tokens, enabled by a custom LongCat Sparse Attention (LSA) mechanism that evolved from DeepSeek's sparse attention design .
• Active parameter range: 33B–56B per token via dynamic 'Zero-Compute Expert' routing, ensuring no wasted computation .
• N-gram Embedding: An additional ~135 billion parameters in the embedding layer, using 2-to-5 character n-grams to improve local context representation without stacking more MoE layers .
• Purpose-built for: Agentic coding — code understanding, generation, and execution in real-world tool-use scenarios .
• Training infrastructure: Trained from scratch on 50,000–60,000 Chinese-made AI ASICs (not Nvidia GPUs), consuming over 35 trillion tokens of Chinese, English, multilingual text, and source code .
• License: Released under the MIT open-source license .
• Availability: Weights and code published on GitHub, Hugging Face, and the official LongCat website; also accessible via API and an online demo .

Performance and Benchmarks

Meituan claims LongCat-2.0 achieves performance comparable to Google's Gemini 3.1 Pro . Before its official unveiling, the model operated anonymously as 'Owl Alpha' on OpenRouter, where it reportedly topped developer rankings for coding benchmarks .

Key benchmark scores published by the LongCat team on X include: Terminal-Bench 2.1: 70.8, SWE-bench Pro: 59.5 (with GPT-5.5 at 58.6 for comparison), SWE-bench Multilingual: 77.3, and FORTE: 73.2 .

Strategic Significance

LongCat-2.0 carries implications far beyond benchmark scores:

• First trillion-parameter model trained end-to-end on domestic Chinese chips, from pre-training through inference and deployment — a milestone that even DeepSeek-V4-pro did not achieve, as it relied on domestic hardware only for inference .
• Demonstrates China's ability to achieve near-frontier AI capability without access to Nvidia's advanced GPUs, which are subject to U.S. export restrictions .
• Signals potential decoupling of Chinese AI development from U.S. hardware supply chains, with implications for the global AI chip market and geopolitical tech competition .
• For Meituan specifically, it positions a food-delivery and local-services giant as a serious contributor to open-source AI foundation models, similar to moves by Alibaba (Qwen) and ByteDance .

Key Architectural Innovations

LongCat-2.0 introduces two notable architectural improvements over its predecessor, LongCat-Flash:

1. LongCat Sparse Attention (LSA): An evolution of DeepSeek's sparse attention mechanism (DSA), LSA addresses latency bottlenecks in the indexer through three independent efficiency optimizations: flow-aware indexing, cross-layer indexing, and hierarchical indexing — designed to accelerate long-context processing without sacrificing model quality .

2. MOPD (Multi-Objective Process Decoding): The model organizes its experts into three specialized groups — Agent, Reasoning, and Interaction — with a gate router that directs each token to the appropriate group based on task type .

Availability and Access

Developers and researchers can access LongCat-2.0 under the permissive MIT license. The model weights, inference code, and documentation are available on GitHub, Hugging Face, and the official LongCat website. An API endpoint and an interactive online demo are also provided .