The 10 November post opened the weights
Meituan's engineering post dated 10 November 2026 says LongCat-2.5 is released under the MIT license, covering commercial inference and further training. The model card adds that the license does not grant Meituan trademarks or patents. The Hugging Face model card states 1.6 trillion total parameters and about 48 billion active per token. It is a native multimodal mixture-of-experts model. The card's chat link is https://longcat.ai/.
The deployment section on that card says the model can be deployed on NPU platforms and points the NPU notes at SGLang-FluentLLM. A laptop is outside that section. This guide stops at the chat site and the hosted API, both of which a single machine can reach.
Weight card: huggingface.co/meituan-longcat/LongCat-2.5
Open-source post: tech.meituan.com/2026/11/10/Meituan-LongCat-2.5-Open-Source.html
NPU notes: github.com/meituan-longcat/SGLang-FluentLLM/blob/longcat-2.5-npu/npu_test/ReadMe.md
Try it: longcat.ai
One key, then one completion
The API quick start asks for an account on the LongCat API platform, then a key created by hand on the API Keys page. The key is shown once. Lose it and the next step is to create another.
The curl sample posts to https://api.longcat.chat/openai/v1/chat/completions. The same page sets the OpenAI SDK base_url to https://api.longcat.chat/openai. That client appends /chat/completions, which lands on https://api.longcat.chat/openai/chat/completions. On 16 Nov 2026 a POST with an invalid key to both URLs returned 401 and invalid_api_key, so both routes answered. Use the documented base_url. The model string is LongCat-2.5. The Anthropic-compatible base is https://api.longcat.chat/anthropic. The key page is on longcat.chat. The docs are on longcat.ai. The quick start does not say whether the two hosts share one login. Create the key on the key page.
On the params page the context length is 1,048,576 tokens, and that total includes input plus output. The max_tokens cap is 262,144, output only. The quick-start sample happens to send max_tokens: 1000, which sits inside that cap. Temperature is fixed: 1.0 when thinking is on, 0.7 when thinking is off. A different value is overwritten, and the call does not fail for that. top_p defaults to 0.95, range (0, 1]. An out-of-range top_p, top_k, repetition_penalty, or max_tokens returns HTTP 400 with invalid_request_error / invalid_parameter.
When a call is rate limited, the documented sample returns HTTP 429, code rate_limit_exceeded, and retry_after 60.
| Path | What to set | Bound on the docs, 16 Nov 2026 |
|---|---|---|
| Chat site | https://longcat.ai/ | Linked from the model card |
| OpenAI POST | base_url https://api.longcat.chat/openai, model LongCat-2.5 | Context 1,048,576. max_tokens cap 262,144 |
| Claude Code | ANTHROPIC_BASE_URL https://api.longcat.chat/anthropic | Both model fields set to LongCat-2.5 |
| Launch price per 1M tokens | USD 0.30 / 0.006 cached / 1.20 | CNY 2 / 0.04 / 8, limited-time |
from openai import OpenAI
client = OpenAI(api_key="YOUR_APP_KEY", base_url="https://api.longcat.chat/openai")
client.chat.completions.create(model="LongCat-2.5", messages=[{"role":"user","content":"Hello"}], max_tokens=1000)Create a key: longcat.chat/platform/api_keys
Quick start: longcat.ai/platform/docs/zh
Launch prices, printed in two currencies
The LongCat-2.5 pricing page, read 16 November 2026, labels the figures a launch discount. The page HTML lists two tables for the same three rows. In yuan per million tokens: uncached input 2, cached input 0.04, output 8. In dollars per million tokens: uncached input 0.30, cached input 0.006, output 1.20. The line under the table says the discount is a limited-time launch price and the bill follows settlement.
The trademark and patent sentence above is the one on that model card.
Which door to open first
Choose the chat site if the next step is the web UI linked from the model card, before any base URL is wired. The card links https://longcat.ai/.
Choose the OpenAI base URL if the code already calls chat.completions. Set base_url to https://api.longcat.chat/openai and the model to LongCat-2.5. Leave temperature unset if the default for the thinking switch is acceptable. A custom temperature is overwritten.
Choose the Claude Code settings file if Claude Code is already installed. The published sample sets ANTHROPIC_AUTH_TOKEN to the LongCat key, ANTHROPIC_BASE_URL to https://api.longcat.chat/anthropic, and both ANTHROPIC_MODEL and ANTHROPIC_SMALL_FAST_MODEL to LongCat-2.5. Copy the remaining keys from that page. Keep the key out of the repo.
Choose the weight repo if an NPU host is actually available and the SGLang-FluentLLM notes are the next read. The card does not describe a laptop install.
Claude Code setup: longcat.ai/platform/docs/zh/claude-code