LongCat-2.5 Open Weights: Meituan’s 1.6T MIT Model, Benchmarks, API Price and Hardware

Meituan, the Chinese food-delivery giant, has released the weights of LongCat-2.5 under the MIT licence. It is a 1.6-trillion-parameter mixture-of-experts model with about 48B active parameters, and it reads images and video as well as text.

The weights landed on Hugging Face on 9 November 2026, a week after Meituan introduced the model. On Meituan’s own tables it trades blows with Claude Opus 5 and DeepSeek V4.1 Flash on agent and coding tests.

This guide covers the specs, every headline benchmark, the API price, how to call it (including from Claude Code), what it really takes to self-host, and the licence fine print.

What LongCat-2.5 Is

LongCat is Meituan’s model family. LongCat-2.0 arrived on 30 June 2026, and the multimodal LongCat-Next research came before this release. LongCat-2.5 merges both ideas into one model.

Its trick is how it sees. Native-resolution images are turned into discrete visual tokens that sit in the same space as text, so the model reasons over a screenshot the way it reasons over code.

• Built for: long agent tasks: browsing, multi-step coding, working through a whole codebase.
• Visual coding: inspect a screenshot or video and build or fix the matching app.
• Hybrid thinking: switch reasoning on for hard tasks, off for cheaper, faster replies.
• Trained on Chinese chips: Meituan says training ran entirely on domestic accelerators.

Specs at a Glance

Note: the 1.6T vs 1.8T gap is unexplained. The extra may be the vision components or F32 tensors. Size your hardware from the actual file total, not the headline.

Benchmarks vs Opus 5, GPT-5.6 Sol and DeepSeek

All numbers below come from Meituan’s model card. LongCat’s own scores are in-house and vendor-reported; rival scores marked by Meituan come from public reports. No independent evaluation had been published when I checked.

How to read these numbers

• Huge jump over 2.0: DeepSWE went from 9.7 to 64.6, and Terminal Bench v4.0 from 3.0 to 38.5.
• Frontier-adjacent, not frontier: on the hardest test (Terminal Bench v4.0) it trails Opus 5 by 15 points. Newer models do better still: the tbench.ai leaderboard lists Claude Opus 5.5 at 64.8 and GPT-6.1 Sol at 58.2 with their own agent harnesses.
• Where it leads: BrowseComp and 3DCodeBench, which fit the browsing-agent and visual-coding pitch.

API Pricing and What a Run Costs

Meituan’s API is the practical way to use LongCat-2.5. Prices below are Meituan’s limited-time discounted rates, and the page warns they can change.

Rival prices are list prices as reported by VentureBeat on 9 November. For the full picture, see our AI API pricing comparison.

Tip: the $0.006 cached input rate is 50x cheaper than uncached. Agents that resend a long system prompt and file context every turn benefit the most.

How to Use the LongCat-2.5 API

LongCat speaks both the OpenAI and Anthropic API formats, so existing SDK code works with a new base URL.

1. Get a key. Sign up on the LongCat API platform at longcat.ai and create an API key.
2. Install the SDK. Run pip install openai.
3. Send a request. Use the code below; set enable_thinking to False for cheaper replies.
4. Add images. Pass an image_url block with a public URL or a base64 data URL. Videos must be public URLs.

from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY",
                base_url="https://api.longcat.chat/openai")

resp = client.chat.completions.create(
    model="LongCat-2.5",
    messages=[{"role": "user", "content": [
        {"type": "text", "text": "Rebuild this page as HTML and CSS."},
        {"type": "image_url", "image_url": {"url": "https://example.com/screenshot.png"}},
    ]}],
    temperature=1.0, top_p=0.95,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)
print(resp.choices[0].message.content)

Meituan’s recommended sampling is temperature 1.0, top_p 0.95 with thinking on. Function calling uses the standard OpenAI tools format.

Image and video limits

• Image formats: PNG, JPG, JPEG, WebP, GIF. No SVG or HEIC.
• Image size: up to 10 MB, 5,120 px per side, 5,242,880 total pixels, 100 images per request.
• Video: URL only, up to 100 MB, under one hour, 10 videos per request.

Use LongCat-2.5 in Claude Code

Because there is an Anthropic-compatible endpoint, you can point Claude Code at LongCat-2.5 with environment variables. This is the standard Claude Code custom-endpoint setup.

export ANTHROPIC_BASE_URL="https://api.longcat.chat/anthropic"
export ANTHROPIC_AUTH_TOKEN="YOUR_LONGCAT_KEY"
export ANTHROPIC_MODEL="LongCat-2.5"
claude

On Windows PowerShell use $env:ANTHROPIC_BASE_URL="..." for each line. For switching between several providers, our Claude Code Router guide shows a cleaner setup.

Warning: your code and prompts go to Meituan’s servers in China. Check your company’s data rules before sending private repositories.

Can You Run It Locally? Hardware Reality

Short answer: not on a PC. Only ~48B parameters fire per token, but all 1.6T must sit in memory.

These are estimates from parameter count (2 bytes per BF16 parameter), scaled from Kimi K3’s published GGUF sizes. They exclude KV cache, which grows with long contexts.

• Official route: Meituan’s SGLang-FluentLLM branch longcat-2.5-npu, written for Chinese NPUs.
• vLLM / SGLang mainline: no official recipe at release. The repo uses custom code, so expect to wait for upstream support.
• llama.cpp / Ollama / LM Studio: no support or GGUF yet.

Note: even 2x RTX 5080 with 128 GB RAM (160 GB total) is several times short. If you want a big open model at home, see running DeepSeek V4.1 Flash locally or our VRAM guide.

LongCat-2.5 vs Other Open-Weight Giants

Sources: model cards and our Kimi K3 VRAM guide and DeepSeek V4.1 Flash guide.

Which should you pick?

• Cheapest strong coder via API: DeepSeek V4.1 Flash still wins on price and DeepSWE.
• Screenshot-to-app and browsing agents: LongCat-2.5 is worth testing; it leads BrowseComp and 3DCodeBench in Meituan’s table.
• Hardest terminal tasks: closed models like Claude Opus 5.5 remain clearly ahead.
• Running at home: neither LongCat nor Kimi; pick DeepSeek V4.1 Flash at 1-bit or a smaller model.

Licence, Limits and Unknowns

• MIT licence: commercial use, modification, fine-tuning and redistribution are allowed if you keep the notice.
• No trademark or patent grant: the card says the licence does not give rights to Meituan trademarks or patents.
• Your responsibility: Meituan says the model is not fully evaluated for every use and asks developers to check accuracy, safety and fairness for high-risk settings.
• Unknowns: API context window, max output, rate limits and how long the discounted price lasts.
• Data location: the hosted API is run by Meituan; self-hosting is the only way to keep data in-house.

Troubleshooting

My Take

LongCat-2.5 is the most interesting kind of open release: not the best model, but a near-frontier model under MIT with genuine visual reasoning. That combination is rare.

The jump from 2.0 is dramatic, and Meituan training it on domestic chips is a story in itself. But all scores are self-reported, and the hardest benchmarks still show a clear gap to Claude and GPT.

For most readers, the API is the real product. At $0.18 for a 300K-token run, it’s cheap enough to test as a visual-coding agent today. Wait for independent benchmarks before betting a production pipeline on it.

FAQ

Is LongCat-2.5 really open source?

The weights are released under the MIT licence, which allows commercial use and modification. Training data and full training code are not released.

How many parameters does LongCat-2.5 have?

About 1.6 trillion in total with roughly 48 billion active per token, per the model card. Hugging Face metadata shows 1.8T.

Can I run LongCat-2.5 on my gaming PC?

No. Even a 4-bit copy needs around 0.8-0.9 TB of memory by my estimate, and there is no GGUF or llama.cpp support yet.

How much does the LongCat-2.5 API cost?

$0.30 per million input tokens, $0.006 cached, and $1.20 per million output tokens, as a limited-time price.

Is LongCat-2.5 better than Claude Opus 5?

Not overall. Meituan’s own numbers show it close on Terminal Bench v2.1 and ahead on BrowseComp, but 15 points behind on Terminal Bench v4.0.

Can LongCat-2.5 understand video?

Yes. It scored 85.2 on VideoMME (vendor-reported), and the API accepts video by public URL, up to 100 MB and under an hour.

Can I use it with Claude Code?

Yes, through the Anthropic-compatible endpoint by setting ANTHROPIC_BASE_URL to https://api.longcat.chat/anthropic and your LongCat key.

Sources

• Hugging Face: meituan-longcat/LongCat-2.5 model card
• LongCat API platform: LongCat-2.5 pricing
• LongCat API docs: image and video understanding
• LongCat blog: LongCat-2.5
• VentureBeat: Meituan releases LongCat-2.5 open weights under MIT
• ai-tldr.dev: LongCat-2.5 release notes