News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnLongCat-2.0

LongCat-2.0

A 1.6 trillion parameter MoE language model from Meituan.

LongCat-2.0 is a large mixture of experts language model developed by Meituan. It features 1.6 trillion total parameters and supports context windows of one million tokens. The model uses custom sparse attention mechanisms and was trained on AI ASIC hardware.

Share on WhatsApp

  • Contains 1.6 trillion total parameters with 48 billion active per token
  • Trained on over 35 trillion tokens using AI ASIC superpods
  • Supports a context window of one million tokens
  • Released under the MIT Licence for open use
  • Uses LongCat Sparse Attention for efficient memory access

What it is

LongCat-2.0 is a large-scale language model built by the Meituan LongCat team. It operates as a mixture of experts architecture, containing 1.6 trillion total parameters while activating approximately 48 billion per token. The system is designed to handle long-horizon tasks, coding, and agentic workflows. It supports a context window of one million tokens, allowing it to process vast amounts of information in a single pass. The model weights are available under an open licence for developers to use and modify.

How it came about

The model represents a significant expansion from earlier versions in the LongCat series. Its development focused on demonstrating that frontier-scale training could occur on alternative hardware platforms rather than standard graphics processors. The team conducted the full pretraining run on AI ASIC superpods, accumulating millions of accelerator-days of compute. This process involved no rollbacks or irrecoverable loss spikes, indicating stable training dynamics at scale.

Inside the model

The architecture introduces LongCat Sparse Attention to manage memory access and scoring efficiency. This method uses streaming-aware indexing to align token selection with hardware capabilities, turning fragmented reads into sequential ones. Cross-layer indexing allows a single indexing pass to serve multiple consecutive layers, reducing computational cost. Hierarchical indexing employs a two-stage scoring scheme to shrink the candidate space. The model also incorporates 135 billion N-gram Embedding parameters to improve parameter utilisation efficiency alongside the sparse expert layers.

What it can do

The model demonstrates strong performance in code understanding, repository-level edits, and automated task execution. It integrates with mainstream harnesses such as Claude Code, OpenClaw, and Hermes to support agentic workflows. Training on hundreds of billions of tokens with one-million-token contexts enhances its capability for long-horizon reasoning. Evaluations compare it against leading proprietary models across agentic, coding, search, and productivity tasks, though specific scores are measured in-house.

How to use it

Model weights are released under the MIT Licence, permitting broad use and modification. Developers can deploy the model on both GPU and NPU platforms using provided cookbooks and libraries like SGLang. A chat interface is available on the official website for direct interaction. The release includes a chat template for encoding messages, supporting both thinking and non-thinking modes for different efficiency needs.

Safety and testing

The developers state that the model has not been comprehensively evaluated for every possible downstream application. Users are advised to assess accuracy, safety, and fairness before deploying it in sensitive or high-risk scenarios. Responsibility for compliance with data protection, privacy, and content safety laws rests with the developers and downstream users. The licence does not grant rights to use Meituan trademarks or patents.

Questions people ask about LongCat-2.0

What is the parameter size of LongCat-2.0?

The model has 1.6 trillion total parameters. However, it uses a mixture of experts design, so only about 48 billion parameters are activated for each token processed. This allows for high capacity with lower computational cost during inference.

Can I use LongCat-2.0 commercially?

Yes, the model weights are released under the MIT Licence. This permissive licence allows for commercial use, modification, and distribution. However, it does not grant rights to use Meituan trademarks or patents associated with the brand.

What hardware is needed to run this model?

The model can be deployed on both GPU and NPU platforms. The developers provide specific deployment guides and cookbooks for SGLang to help users set up the model on these different types of accelerator hardware.

How long is the context window?

LongCat-2.0 supports a context length of one million tokens. This was achieved by training on hundreds of billions of tokens with long contexts and implementing specialized sparse attention mechanisms to handle the data efficiently.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on LongCat-2.0 yet, so the facts on this page come from the sources listed above: huggingface.co/meituan-longcat/LongCat-2.0. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.