Try LongCat 2.0 Online
Free AI demo of Meituan's 1.6T-parameter Mixture-of-Experts model. 33B–56B dynamic activation, 1M context window, trained entirely on domestic AI chips.
🔒 Verification helps us keep this free demo available for everyone.
Why Try LongCat 2.0?
A production-grade open model built from the ground up on domestic AI infrastructure.
1.6T MoE Architecture
Massive 1.6 trillion parameter model with only 33B–56B dynamically activated per token, delivering high performance with efficient inference.
1 Million Token Context
LongCat Sparse Attention enables ultra-long context understanding. Process entire books, codebases, or long documents in one pass.
ScMoE & Dynamic Activation
Scalable MoE overlaps communication with computation. Empty experts route simple tokens to 33B while complex tokens use 56B.
100% Domestic Chips
Trained on over 50,000 non-CUDA domestic AI accelerators with ~60GB memory each. A milestone for independent AI infrastructure.
Agent-Ready
Strong performance on coding, tool use, and multi-step agent benchmarks. Try it for coding assistants, SQL agents, and automation.
Open Source MIT License
Full weights available on Hugging Face, including FP8 and INT8 quantized versions. Free for commercial and research use.
LongCat 2.0 Benchmarks
Honest comparison against first-tier closed-source models on agent-oriented tasks.
| Model | Parameters | Agent Average | Context Window |
|---|---|---|---|
| Claude 4 | Unknown | 92.5 | 200K |
| GPT-4o | Unknown | 90.8 | 128K |
| Gemini 3.1 Pro | Unknown | 86.4 | 1M |
| LongCat 2.0 | 1.6T MoE (33–56B active) | 85.2 | 1M |
Scores are illustrative composites based on agent-oriented benchmarks. Visit the official LongCat 2.0 GitHub for full evaluation details.
Three Ways to Use LongCat 2.0
Try it online, integrate via API, or run it on your own hardware.
Try Online Demo
Chat with LongCat 2.0 directly on this page. No signup, no setup — just complete the verification and start asking.
Try NowUse the API
Get an API key from longcat.ai. Pricing starts at $9.9 for 50 million tokens, with 100% free cached tokens.
Self-Host
Download weights from Hugging Face. Available in full precision, FP8, and INT8 quantized formats.
Frequently Asked Questions
Everything you need to know about trying LongCat 2.0 online.
LongCat 2.0 is an open-source large language model developed by Meituan. It uses a 1.6 trillion parameter Mixture-of-Experts (MoE) architecture with 33B–56B dynamically activated parameters and supports up to 1 million tokens of context.
Yes. You can try LongCat 2.0 online for free on trylongcat.com or on the official platform longcat.ai. The model is also open-source under the MIT license, so you can self-host it at no cost.
On agent-oriented benchmarks, LongCat 2.0 ranks behind Claude and GPT-4o but is competitive with Gemini 3.1 Pro. Its key differentiator is that it was trained entirely on domestic AI accelerators, not NVIDIA GPUs.
LongCat 2.0 was trained entirely on domestic AI ASIC accelerators, peaking at over 50,000 cards. The training cluster uses non-CUDA accelerators with approximately 60 GB of available device memory per card.
You can download LongCat 2.0 weights from the official Hugging Face repository: huggingface.co/meituan-longcat/LongCat-2.0. Quantized FP8 and INT8 versions are also available.
No. trylongcat.com is an independent community demo site. It is not officially affiliated with Meituan or the LongCat team. All trademarks belong to their respective owners.
Ready to Try LongCat 2.0?
Start chatting for free. No credit card required. Experience Meituan's open-source 1.6T MoE model today.
Start Free Demo