LongCat-2.0
Meituan's open-source trillion-parameter model for coding and agentic tasks
Overview
LongCat-2.0 is Meituan's follow-up open-source model, and the headline number is 1.6 trillion total parameters in a mixture-of-experts setup that only activates around 48 billion per token, plus a native 1 million token context window. It was trained on tens of trillions of tokens with heavy emphasis on coding and agentic tasks, and Meituan has been positioning it as competitive with other large open models on repository-level coding benchmarks. The weights are released under the MIT license, so anyone can download, self-host, or fine-tune it without paying a licensing fee.
This is not a consumer product with a pricing page. There's a free hosted chat demo at longcat.ai where you can try it without setup, and the model itself is available on Hugging Face for anyone with the hardware or cloud budget to run it. That makes it a fit for developers, researchers, and companies who want an open alternative to closed models for coding-heavy workloads, not for someone looking for a polished chat assistant with a support team behind it. If you need something to just work out of the box with no infrastructure decisions, this isn't that.
Key Features
1.6T Parameters, MoE: Large mixture-of-experts architecture with roughly 48B parameters activated per token, keeping inference cost lower than a dense model of similar size.
Native 1M Context: Trained on long-context data, useful for large codebases or long documents.
LongCat Sparse Attention: A custom attention mechanism aimed at making long-context inference more efficient.
Agentic Coding Focus: Tuned for repository-level edits and multi-step coding tasks, with integrations into tools like Claude Code.
MIT License: Fully open weights, free to download, self-host, and fine-tune.
Free Hosted Demo: Chat with the model directly at longcat.ai without setting up your own infrastructure.
Pricing
Starting price
Free (open-weight model under MIT license, no paid consumer tier)
Model Weights: Free, MIT license, download from Hugging Face or ModelScope and self-host.
Hosted Demo: Free to chat with at longcat.ai, no published pricing for higher-volume API access at time of writing.
Third-Party API Providers: Some inference platforms host LongCat-2.0 with their own usage-based pricing; check each provider directly.
Disclaimer: pricing and hosting options may change, confirm on Meituan's official LongCat channels before relying on any figure.
Pros
Genuinely Open: MIT license means no usage fees and full freedom to self-host or fine-tune.
Strong on Coding: Built and benchmarked specifically for agentic coding and repository-scale edits.
Huge Context Window: Native 1M-token context is genuinely useful for large codebases.
Free Demo Access: You can try it at longcat.ai before committing to self-hosting.
Cons
Not Consumer-Ready: No polished app, no customer support, no simple subscription; self-hosting requires serious infrastructure and expertise.
Enormous Compute Needs: A 1.6T parameter model is out of reach for most individuals to run locally, even quantized.
Sparse Independent Reviews: Too new for a large body of third-party benchmarking outside Meituan's own claims.
Data Origin Questions: As with most large open models, training data provenance isn't fully disclosed.
What Makes It Unique
Open Weights at Trillion-Parameter Scale: Very few labs release a model this large under a fully permissive MIT license instead of a restrictive or closed one.
Kay Score
7
/ 10
Tool Information
Pricing
Free (open-weight model under MIT license, no paid consumer tier)
Category
Coding & Development
Platform
Web / iOS / Android
Last Updated
