Home / Guides

LongCat 2.0 on Hugging Face: Repos, Files & Download Methods Explained

Published September 7, 2026 · by the getlongcat.com team

The official home of LongCat 2.0 weights is the meituan-longcat organization on Hugging Face. This guide explains what is actually in those repos — because a trillion-parameter checkpoint is not a single file you right-click and save — and compares the four practical ways to pull the weights down. For a channel-by-channel comparison including the China mirror, see our LongCat 2.0 download hub.

1. The three model repos

The organization publishes one repo per precision variant:

All three are the same 1.6T-parameter MoE with 33–56B active parameters per token — quantization changes storage precision, not architecture. The model card on each repo page is the authoritative spec sheet: parameter counts, context length, recommended serving flags and the MIT license text.

2. What the files inside actually are

Open the "Files" tab on any of the three repos and you will find the same anatomy:

• *.safetensors shards — the weights themselves, split into dozens of numbered shards because no sane filesystem operation wants a single multi-terabyte object. The shards plus the index file together are the model; a partial download is not a smaller model, it is a broken one.
• model.safetensors.index.json — the map from tensor name to shard file. Loaders (transformers, vLLM, SGLang) read this first to know which shard holds which expert.
• config.json — architecture hyperparameters: layer counts, expert counts, hidden sizes, rope settings. Serving stacks refuse to load without it.
• Tokenizer files — tokenizer.json / tokenizer.model and friends. Small, easy to overlook, fatal to forget.
• LICENSE / README — the MIT license text and the model card.

The practical consequence: always download the complete repo, not individual shards, unless you know exactly which tensors you need (almost nobody does).

3. Four ways to download

Method A — Hub CLI (recommended)

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

The CLI resumes interrupted transfers and parallelizes across files — with hundreds of gigabytes in flight, both matter. This is the method the official docs assume.

Method B — Python API

from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="meituan-longcat/LongCat-2.0-FP8",
    local_dir="./longcat-2-fp8",
    max_workers=8,
)
print(path)

Use this when the download is part of a provisioning script rather than a one-off manual step.

Method C — Selective files

huggingface-cli download meituan-longcat/LongCat-2.0-FP8 \
  config.json tokenizer.json model.safetensors.index.json \
  --local-dir ./longcat-2-fp8-meta

Passing file names pulls only those files. Useful for inspecting the config or index on a laptop before committing a server to the full transfer.

Method D — Mirror endpoints

From networks where huggingface.co is slow or unreachable, point the same tooling at a mirror endpoint:

HF_ENDPOINT=https://hf-mirror.com \
  huggingface-cli download meituan-longcat/LongCat-2.0-FP8 --local-dir ./longcat-2-fp8

For mainland China specifically, the first-party option is the official ModelScope mirror published by the same organization — preferable to third-party mirrors because the files come from Meituan's own account. Commands are in the download hub.

4. After the download: loading the weights

Hugging Face transformers can load the checkpoint for inspection, but production inference at this scale belongs to vLLM or SGLang — the official GitHub repo documents the launch flags per checkpoint, and our getting started guide walks through a working serving setup. If this all sounds like more infrastructure than you signed up for, the managed route in our API quickstart skips the download entirely.

5. A note on unofficial re-uploads

Search Hugging Face for "LongCat" and you will find community re-uploads: GGUF conversions, further quantizations, merges. Some are genuinely useful — a GGUF quant is the only way to experiment with a slice of this model on consumer hardware. But they are not the released model: quality regressions and conversion bugs are the uploader's, not Meituan's. For anything you intend to ship, start from the official meituan-longcat repos and treat community conversions as experiments.