LongCat-Flash-Lite-Sparse

LongCat-Flash-Lite-Sparse

Mixture-of-experts text model with 69B parameters, about 3B active per token, sparse attention and a 1M-token context. For teams hosting long-context models for coding and tool-use agents.

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-Flash-Lite-Sparse is a text model from the LongCat team with 69 billion parameters in total and about 3 billion active for each token. It does not show a thinking phase. It swaps the dense attention of its predecessor for a sparse attention scheme, and it supports a context window of up to 1 million tokens.

Who it suits: It suits teams that run their own large language model and care about long inputs and agent work, such as coding and tool use.

Industry | Who uses it | How they use it | Example
Software | Platform engineers | Host a long-context model for coding agents | Working through a large repository in one prompt
Research | Model researchers | Study sparse attention and indexing methods | Comparing results with the dense predecessor
Enterprise | AI teams | Run tool-using agents on a self-hosted model | Chaining search and tool calls in an agent loop

You can skip it if you want a small model for a laptop, since the weights are over 100 GB in BF16.

Where to use it: It is meant for server hardware that can hold a 69B model. The download step works on Windows, macOS or Linux with Python.

• A multi-GPU server for serving the weights
• Agent frameworks that call a self-hosted model
• Windows, macOS or Linux, for fetching the files

Use case | Who | What they get back
Agentic coding | Developers | Better results on coding benchmarks than the dense predecessor
Long documents | Analysts | Room for inputs up to 1 million tokens
Tool use and search | Agent builders | Stronger tool-use scores in the card's tests

How to set it up: You need Python, the Hugging Face tool and a lot of storage. Hugging Face lists about 138 GB in BF16, or an estimated 34.6 GB at 4-bit. The card's own pages cover how to run it.

• Python and the Hugging Face tool
• About 138 GB of disk for the BF16 weights
• A serving stack that can handle a large sparse-attention model

1. Install the Hugging Face download tool.
2. Download the repository files to a drive with enough room.
3. Follow the model card to serve it.

pip install -U huggingface_hub
hf download meituan-longcat/LongCat-Flash-Lite-Sparse

How to use it: After the weights are served, send it text prompts like any chat model. It is built for long inputs and agent tasks, so give it whole documents or multi-step tool jobs. The visible part of the card gives no run command, so follow its own instructions.

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

Example: You give a served copy a bug report and a large codebase pasted in as one long prompt. It reads the whole input and answers with a proposed fix, as an agentic coding task would need.

For newcomers and experts: If you are new, this is a big model meant for servers, not a file you open on a laptop. Read the card first to see what a full setup needs.

If you know the field, the sparse attention builds on DeepSeek Sparse Attention with three indexing ideas. Streaming-aware indexing keeps a fixed sink and local window. Cross-layer indexing reuses one pass across layers. Hierarchical indexing scores coarse blocks first and can be switched on at inference time without more training.

Long input up to 1 million tokens
Sparse attention picks relevant tokens
Mixture-of-experts layers, 3B active
Text reply or agent action

Good to know: Turning on hierarchical indexing trades a small drop on some coding scores for lower indexing cost.

• License: MIT
• Cost to run: no API key, account or paid service is named, but you need hardware for a 69B model
• Data it reaches: files downloaded from Hugging Face, plus whatever your agent setup lets it read
• Not for: laptops, or tasks that need a visible thinking mode, since it is a non-thinking model

Title: SWE-Bench Verified accuracy
Dense predecessor: 54.40
Sparse without HI: 68.20
Sparse with HI: 65.20

On Hugging Face, 13 November 2026

• Task: text generation
• Library: LongCat-Flash-Lite-Sparse
• Parameters: 69.1B
• Weights: about 138 GB in BF16; about 34.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,200
• Likes: 92
• License: mit
• Quantized versions: 3
• Fine-tunes: 3
• Tags: longcat-flash-lite-sparse, safetensors, transformers, text-generation, conversational

Links

• Model: https://huggingface.co/meituan-longcat/LongCat-Flash-Lite-Sparse
• Files: https://huggingface.co/meituan-longcat/LongCat-Flash-Lite-Sparse/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download meituan-longcat/LongCat-Flash-Lite-Sparse

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