Needle 3

Needle 3

A small on-device model from Cactus Compute for tool calling, structured extraction and text embeddings, used through the cactus Python package. For developers building local agents.

Made by Cactus-Compute 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: Needle 3 is a model from Cactus Compute that runs on the device. The card says it is made for three jobs: calling tools, pulling structured data out of text, and making text embeddings.

You use it through the cactus-compute Python package. This repository provides the model files needle3.cact and checkpoints/needle3.safetensors. The package downloads and keeps the engine and weights on first use.

Who it suits: Developers who want an agent that picks and calls tools on a phone, laptop or other device without a remote server.

Industry | Who uses it | How they use it | Example
Software | App developers | Let a local agent call functions | Look up the weather for a city
Data work | Analysts | Pull structured fields out of text | Turn a message into named fields
Search | Product teams | Make embeddings on the device | Compare short texts locally

Skip it if you want a general chat model or long writing, since the card only describes tool calling, extraction and embeddings.

Where to use it:

• Python programs that import the cactus package
• Consumer devices, which the Cactus banner names as the target
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Tool calling | Agent builders | A call to one of your Python functions with the right arguments
Structured extraction | Data teams | Fields taken out of free text
Embeddings | Search builders | Vectors for text made on the device

How to set it up:

• Python and pip
• The cactus-compute package
• A device the Cactus runtime supports, as listed in the Cactus repository

1. Install the package.
2. Write a Python function and mark it as a tool.
3. Start the agent with your tools and run a request.

pip install -U cactus-compute

How to use it: Mark a function with the tool decorator, give it a short description in a docstring, and pass it to the agent. Then call run with a plain question.

import cactus

@cactus.tool
def get_weather(city: str):
    "Get the current weather for a city."
    return {"city": city, "temp_c": 27, "sky": "clear"}

with cactus.Cactus(tools=[get_weather]) as agent:
    print(agent.run("What's the weather in Lagos?")["results"])

Part | What it does
cactus.tool | Marks a function so the agent can call it
cactus.Cactus | Starts the agent with the tools you give it
agent.run | Sends a request and returns the results

Example: You ask the agent for the weather in Lagos. It chooses the get_weather tool, passes the city name, and returns the dictionary your function produced.

For newcomers and experts: If you are new to tool calling, copy the weather example and change the function to something of your own. Keep each tool small, with a clear description.

If you know the field, the card points to separate guides for the Python API, designing tools, structured extraction and fine-tuning. This page does not repeat benchmark numbers, so check them on the makers' site.

Install the cactus-compute package
Write and mark your tool
Start the agent with the tool
Run a request and read the results

Good to know:

• The first run downloads and caches the engine and the weights
• The banner on the card describes it as under 50 MB, but the card gives no exact file size
• Hugging Face lists 3 quantized and 8 fine-tuned versions made by other people
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the files you download and the requests and tool results you pass through it
• Not for: general chat or long writing

On Hugging Face, 13 November 2026

• Task: text generation
• Library: cactus-compute
• Downloads in the last 30 days: 147,784
• Likes: 411
• License: apache-2.0
• Quantized versions: 3
• Fine-tunes: 8
• Tags: cactus-compute, needle, tool-calling, function-calling, on-device, edge, quantization, webassembly, text-generation

Links

• Model: https://huggingface.co/Cactus-Compute/needle3
• Files: https://huggingface.co/Cactus-Compute/needle3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Cactus-Compute/needle3

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