Home / Guides

Getting Started with LongCat 2.0: Download and Deploy

Published August 2, 2026 · by the getlongcat.com team

LongCat 2.0 is Meituan's open-source Mixture-of-Experts language model: 1.6 trillion total parameters with roughly 33–56B active per token, a context window of up to 1M tokens, and an MIT license that permits commercial use. This guide walks through the practical first steps — choosing a checkpoint, downloading the weights, and running inference.

1. Choose your checkpoint

Three official variants are published on Hugging Face under the meituan-longcat organization:

• LongCat-2.0 (full precision) — the reference checkpoint. Maximum quality, maximum hardware: plan for a multi-GPU server with well over a terabyte of aggregate VRAM.
• LongCat-2.0-FP8 — 8-bit floating point. Roughly half the memory of the full model with minimal quality loss on supported hardware.
• LongCat-2.0-INT8 — 8-bit integer quantization. The lightest official option; useful when VRAM is the binding constraint.

If you just want to evaluate the model rather than self-host it, skip the download entirely: use the official API platform, or try it in the browser at trylongcat.com.

2. Download the weights

The standard way is the Hugging Face Hub CLI:

pip install -U "huggingface_hub[cli]"
huggingface-cli download meituan-longcat/LongCat-2.0-FP8 \
  --local-dir ./longcat-2-fp8

For downloads from mainland China, the same weights are mirrored on ModelScope, which is usually much faster there:

pip install modelscope
modelscope download --model meituan-longcat/LongCat-2.0-FP8

Either way, verify file integrity against the official model card before deployment — the checkpoints are large, and interrupted transfers are common.

3. Plan your hardware

Be realistic: a 1.6T-parameter MoE is server-class infrastructure even though only 33–56B parameters are active per token. All expert weights must fit in memory. As a rule of thumb, the full-precision checkpoint needs on the order of 3 TB of VRAM across multiple GPUs, while FP8 brings that down by roughly half. The official repo also maintains the SGLang-FluentLLM NPU branch for domestic AI accelerators, which matters if you are deploying on non-NVIDIA hardware.

4. Serve with vLLM or SGLang

The official GitHub repository documents serving with vLLM and SGLang. A typical vLLM launch looks like:

vllm serve ./longcat-2-fp8 \
  --tensor-parallel-size 8 \
  --max-model-len 131072

Start with a shorter --max-model-len and scale up once the service is stable — long-context KV cache is the first thing that blows up memory. Check the repo README for the exact flags recommended for each checkpoint, as they change between releases.

5. First request

Both vLLM and SGLang expose an OpenAI-compatible endpoint, so your first smoke test is a standard chat completion against localhost:8000/v1/chat/completions. If you would rather not self-host at all, the managed route is covered in our LongCat 2.0 API quickstart.