Home / Download

LongCat 2.0 Download: Weights, Checkpoints & Mirrors

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

LongCat 2.0 — Meituan's open-source 1.6T-parameter Mixture-of-Experts model with 33–56B active parameters, up to 1M tokens of context and an MIT license — is distributed through a small set of official channels. This page aggregates every legitimate download source, explains which checkpoint fits which hardware, and shows how to verify what you downloaded. If you only want to use the model rather than self-host it, you do not need to download anything at all — skip to the no-download options.

1. Pick a checkpoint

Three official weight variants are published under the meituan-longcat organization. They differ only in numeric precision — same architecture, same training run:

Even the quantized variants are server-class: every expert weight must fit in memory even though only 33–56B parameters activate per token. The hardware planning section of our getting started guide goes deeper, including the SGLang NPU branch for domestic accelerators.

2. Official download channels

Hugging Face (primary)

The canonical weights live on Hugging Face. The Hub CLI handles resume and parallelism for multi-hundred-gigabyte transfers:

pip install -U "huggingface_hub[cli]"

# Full precision
huggingface-cli download meituan-longcat/LongCat-2.0 --local-dir ./longcat-2-full

# FP8 (recommended for serving)
huggingface-cli download meituan-longcat/LongCat-2.0-FP8 --local-dir ./longcat-2-fp8

# INT8
huggingface-cli download meituan-longcat/LongCat-2.0-INT8 --local-dir ./longcat-2-int8

Direct links: full · FP8 · INT8. Each repo page carries the model card with the official file inventory.

ModelScope (China mirror)

For downloads from mainland China, the same checkpoints are mirrored on ModelScope by the same official organization — typically far faster than crossing the Pacific:

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

GitHub (source code, not weights)

The GitHub repository holds the inference code, model card and serving examples for vLLM and SGLang — but not the weight files themselves. Clone it alongside the weights:

git clone https://github.com/meituan-longcat/LongCat-2.0.git

3. Verify your download

Checkpoint transfers are large enough that silent truncation is a real failure mode. Before loading anything:

1. Compare the file inventory — count and names of the .safetensors shards, plus config.json, tokenizer files and the license file, against the official model card listing.
2. Check sizes — a shard that is even a few bytes short will fail to load or, worse, load with corrupted tensors.
3. Load a smoke test — run one short generation on a single GPU subset before standing up the full serving stack. A failed safetensors header check at this stage costs minutes; discovering it mid-deployment costs hours.
4. Trust only the official organization — weights published by accounts other than meituan-longcat (re-uploads, "optimized" repacks) cannot be assumed to match the released model.

4. No-download alternatives

Self-hosting a 1.6T MoE is a serious commitment. Most people evaluating LongCat 2.0 should start with one of these instead:

• Official API — the managed platform at longcat.chat exposes an OpenAI-compatible endpoint; our API quickstart gets you a first call in about five minutes.
• OpenRouter — LongCat 2.0 is listed on OpenRouter, convenient if you want one key for side-by-side benchmarking against other models.
• Browser demo — trylongcat.com runs the model online with zero setup.

5. After the download

Weights on disk are step one. The full path — checkpoint choice, hardware planning, vLLM/SGLang serving flags and the first request — is in Getting Started with LongCat 2.0: Download and Deploy. If you are deciding between API and self-hosting on cost, the comparison in our API quickstart lays out when each route wins.