LongCat-2.5

LongCat-2.5

Meituan LongCat's multimodal mixture of experts model for long agent tasks and coding, with about 1.6 trillion total parameters and 48 billion active per token. Suits labs and teams with large GPU clusters.

Made by meituan-longcat on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: LongCat-2.5 is a model from meituan-longcat that reads images and text. By the card's count it is a mixture of experts design with about 1.6 trillion parameters in total and roughly 48 billion active for each token. It follows LongCat-2.0 and is aimed at big code projects, building whole projects, design work and 3D jobs.

Who it suits: Labs and companies with large GPU clusters that want an open multimodal model for agent style coding, and researchers who study long running agents.

Industry | Who uses it | How they use it | Example
Software | Engineering teams | Run an agent that works across a large codebase | Ask it to build a full project from a brief
Design | Creative teams | Produce design and 3D work from instructions | Turn a written brief into a page design
Research | Model researchers | Study its training approach for long agent tasks | Read the evaluation tables in the card
Infrastructure | Serving teams | Test how to host a huge model | Plan storage for the weights

Skip it if you only have a single computer, because the weights are an estimated 3,560 GB in BF16 and 890 GB at 4 bit.

Where to use it:

• A multi GPU cluster with enough memory for the weights
• Agent and coding tools that you connect to a copy you serve yourself
• Your own storage, filled with the Hugging Face tools

Use case | Who | What they get back
Agent style coding | Engineers | Code changes planned and made across a large codebase
Image and text questions | Teams | Answers that use both a picture and your words
Creative and 3D work | Designers | Design and 3D output from prompts

How to set it up:

• Python and the Hugging Face tools
• Storage for about 3,560 GB in BF16, or an estimated 890 GB at 4 bit
• A large GPU cluster

1. Install the Hugging Face hub package.
2. Download the repository files.
3. Serve them with your own inference setup, since the card text here lists no command.

pip install -U huggingface_hub
hf download meituan-longcat/LongCat-2.5

How to use it: Connect your tools to a served copy and send it text, images or both. The card text here shows no client command, so follow the serving notes on the model page.

Example: Ask it to build a small project from a written brief, with a picture of the wanted design attached. It should read both, plan the steps and write the code across several files.

For newcomers and experts: If you are new to this, know that this model is far too large for a home computer. Read the model page first and try a smaller model for your own first experiments.

If you know the field, the card says visual features at native resolution are turned into discrete tokens and aligned with text in one space. Training uses long horizon agent reinforcement learning, several teacher models and tree attention. The card reports 86 percent more throughput per device than a DDP baseline in long sequence fine tuning.

Images and text go in
Visual features become discrete tokens
The mixture of experts model runs
A reply or agent action comes out

Title: Estimated weight size
BF16: 3560 GB
4 bit: 890 GB

Good to know:

• The card says 1.6 trillion total parameters, while Hugging Face's own count shows 1.8T.
• Scores marked with an asterisk in the card come from public reports, and the rest are in house and indicative only.
• The card text shown here has no run steps, so check the model page before you plan hardware.
• License: mit.
• Cost to run: none named, but the hardware needs are huge.
• Data it reaches: the prompts and images you give it. The source names no account or outside service.
• Not for: the source names none.

On Hugging Face, 13 November 2026

• Task: image text to text
• Parameters: 1.8T
• Weights: about 3,560 GB in BF16; about 890 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 710
• Likes: 186
• License: mit
• Tags: safetensors, pro_v, multimodal, image-text-to-text, text-generation, mixture-of-experts, agentic, conversational

Links

• Model: https://huggingface.co/meituan-longcat/LongCat-2.5
• Files: https://huggingface.co/meituan-longcat/LongCat-2.5/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download meituan-longcat/LongCat-2.5

Then follow the usage in its model card. Nothing in this kit runs by itself.