LongCat-2.5-W4A4

LongCat-2.5-W4A4

Multimodal mixture-of-experts model for long agentic tasks such as codebase work, full project builds and creative design, in a compressed W4A4 build. For teams with large GPU capacity.

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 multimodal mixture-of-experts model from Meituan's LongCat team. It takes text and pictures as input and targets long agent runs: developing in large codebases, building whole projects, design work and 3D tasks. The card gives 1.6 trillion total parameters, with about 48 billion active per token. This repository is the W4A4 build.

Who it suits: It suits teams with serious GPU capacity who want an open model for coding agents that also understands images.

Industry | Who uses it | How they use it | Example
Software | Platform teams | Run a coding agent on a self-hosted model | Building a full project from a written brief
Design | Creative tool builders | Combine image understanding with generation of designs | Turning a screenshot into a page layout
Research | AI labs | Study long-horizon agent training and serving | Comparing it with other open models

You can skip it if you have a single consumer GPU, since even the compressed weights are hundreds of gigabytes.

Where to use it: It is meant for server-class setups that can host a huge model and serve it to agents or apps.

• A multi-GPU server or cluster
• Coding agent setups that need image understanding
• Windows, macOS or Linux machines for the download step

Use case | Who | What they get back
Codebase development | Engineering teams | Agentic coding help across a repository
Full project builds | Agent developers | Multi-step work carried through to a finished project
Image-aware tasks | Designers | Replies that use both pictures and text

How to set it up: You need Python for the download tool and a lot of storage. Hugging Face estimates about 247 GB at 4-bit and about 990 GB in BF16. The card points to a technical blog for full details.

• Python and the Hugging Face tool
• Several hundred gigabytes of free disk
• A serving stack able to host a huge model

1. Install the Hugging Face download tool.
2. Download the repository files to a drive with plenty of space.
3. Follow the model card to serve it with your own stack.

pip install -U huggingface_hub
hf download meituan-longcat/LongCat-2.5-W4A4

How to use it: Once the weights are served, send prompts that mix text and images, and give the model long tasks such as building or fixing a project. The visible part of the card does not list a run command, so use the card's own instructions.

Tool | What it does
hf download | Fetches the model files to your computer

Example: You hand a served copy a screenshot of an app and ask for a working version of it. The model reads the picture, plans the project in steps, and writes code across files as an agent would.

For newcomers and experts: If you are new, this is a huge model that most people reach through a hosted service, not a laptop. Start by reading the model card and the project's technical blog to see what a full setup needs.

If you know the field, the card describes native-resolution visual features turned into discrete tokens, long-horizon reinforcement learning, and multi-teacher distillation. It reports 86 percent higher per-device throughput than a DDP baseline on long-sequence fine-tuning, and about double the decode KV capacity on a tested workload.

Text and image inputs
Visual features become discrete tokens
Mixture-of-experts model reads both
Agentic reply or code output

Good to know: Scores on the card marked with an asterisk come from public reports, and the unmarked ones were measured in-house and are only indicative.

• License: MIT
• Cost to run: no API key, account or paid service is named, but the hardware needs are huge
• Data it reaches: files downloaded from Hugging Face, plus whatever your agent setup lets it read
• Not for: small machines or quick laptop trials

On Hugging Face, 13 November 2026

• Task: image text to text
• Parameters: 495B
• Weights: about 990 GB in BF16; about 247 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 240
• Likes: 3
• 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-W4A4
• Files: https://huggingface.co/meituan-longcat/LongCat-2.5-W4A4/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-W4A4

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