LongCat-2.0: Open-Source 1M Context That Beats GPT-5.5

# The Numbers

# Why It Matters

For 2 months, a model called "Owl Alpha" topped OpenRouter's developer charts. Most-used coding model globally. Nobody knew who built it.

It was Meituan - a Chinese food delivery company. They just unmasked it as LongCat-2.0 and open-sourced everything under MIT.

# How to Get It

Three paths: use the API, try the official chat, or download the full weights and self-host.

01 Download the Weights

Two variants available. FP8 is recommended for most deployments (half the VRAM, negligible quality loss).

# Install the CLI
pip install huggingface_hub

# Download FP8 weights (~900GB)
huggingface-cli download meituan-longcat/LongCat-2.0-FP8 \
    --local-dir ./LongCat-2.0-FP8

02 Use via API (Quickest Path)

OpenRouter exposes LongCat-2.0 as an OpenAI-compatible endpoint. Swap your model string and you are done.

import openai

client = openai.OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="your-openrouter-key"
)

response = client.chat.completions.create(
    model="meituan/longcat-2.0",
    messages=[{"role": "user", "content": "..."}]
)

print(response.choices[0].message.content)

03 Self-Host with SGLang

Meituan recommends 16x H20 GPUs (2 nodes) with SGLang. You need to compile a custom sgl-kernel first.

# Install SGLang
pip install uv
uv pip install "sglang[all]>=0.5.2.rc0"

# Compile sgl-kernel (required for LongCat-2.0)
cd sgl-kernel
python3 -m uv build --wheel --color=always --no-build-isolation \
    -Ccmake.define.SGL_KERNEL_ENABLE_SM90A=1 \
    -Ccmake.define.CMAKE_POLICY_VERSION_MINIMUM=3.5 \
    -Cbuild-dir=build .
pip3 install dist/sgl_kernel-*.whl --force-reinstall

python -m sglang.launch_server \
  --model meituan-longcat/LongCat-2.0-FP8 \
  --trust-remote-code \
  --host 0.0.0.0 --port 13423 \
  --tp 16 --ep 16 \
  --max-running-requests 64 \
  --mem-fraction-static 0.92 \
  --chunked-prefill-size 2048 \
  --nsa-prefill-backend fa3 \
  --kv-cache-dtype bfloat16 \
  --nnodes 2 --node-rank 0 \
  --dist-init-addr $MASTER_IP:20000

# The China Chip Angle

LongCat-2.0 trained on a cluster of 50,000 domestic Chinese ASICs. Zero Nvidia GPUs. This is the first near-frontier model to prove that non-Nvidia hardware can reach this level.

The geopolitical implications are significant. US export controls were designed to prevent exactly this. But Meituan just shipped a model that outperforms GPT-5.5 on coding tasks without a single restricted chip.

# Agent Harness Integration

LongCat-2.0 is already integrated into the major coding agent harnesses. Point your tool at the OpenRouter endpoint and it works out of the box.

# Set OpenRouter as your provider
export OPENAI_API_KEY=your-openrouter-key
export OPENAI_BASE_URL=https://openrouter.ai/api/v1

# Launch with LongCat-2.0
claude --model meituan/longcat-2.0