Needle 2

Needle 2

A 45M-parameter open model for tool calling, device use and structured extraction, packed into one 14 MB binary that runs in about 28 MB of RAM. For developers who need function calling on phones, boards and browsers.

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 2 is a small open model, 45M parameters in size, built to call tools, operate devices and pull out structured data. The whole model is one 14 MB binary that needs about 28 MB of RAM for a full session. Text goes in and structured JSON comes out, and a grammar built from your schemas constrains every token.

Who it suits: Developers who want function calling on small hardware, such as phones, single-board computers, VR headsets and microcontrollers. It also suits anyone who needs structured data pulled from text without a big model or a network call.

Industry | Who uses it | How they use it | Example
Mobile apps | App developers | Run tool calling on the phone itself | Let a user say what they want and call the matching function
Embedded devices | Hardware makers | Put a function-calling model on a small board | Control a device from a spoken or typed request
Smart home | Home automation developers | Turn requests into structured commands | Map a sentence to a thermostat setting
Agents | Agent builders | Route among a large tool catalog cheaply | Declare many tools and let retrieval pick the top few

You can skip it if you need open-ended writing or long reasoning. The card describes tool calling, device use and structured extraction.

Where to use it:

• Python on Windows, macOS and Linux, through the cactus-needle package
• Chip families the card names: ARM64, x86-64, ARMv7 and RISC-V, plus WebAssembly in the browser
• Phones, Raspberry Pi boards, VR devices and microcontrollers such as the ESP32-P4

Use case | Who | What they get back
Tool calling | Developers | A structured call with typed arguments
Structured extraction | Data teams | JSON that follows your schema
Device control | Hardware makers | Actions chosen from the tools you declare

How to set it up: You need Python and one pip package.

• Python on Windows, macOS or Linux
• The cactus-needle package, which holds the model and engine
• No other downloads or network access at run time, per the card

1. Install Python if you do not have it.
2. Install the package with pip.
3. Describe your tools as Python functions.
4. Create the agent and run a request.

pip install cactus-needle

How to use it: Decorate a Python function with the tool decorator. Its signature gives the argument types and its docstring is the tool description. Then create the agent and call run. Needle chooses which function to call, runs it with your code, passes the output back to the model and then gives you the closing answer.

import needle

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

agent = needle.Needle(tools=[get_weather])
print(agent.run("what's it like in Lagos right now?")["reasoning"])

For more control, describe each argument in a Google-style Args block, give defaults for optional arguments, and use a Literal for a fixed set of choices the model cannot step outside of.

from typing import Literal

@needle.tool
def set_thermostat(temperature: int, mode: Literal["heat", "cool", "auto"] = "auto"):
    """Set the thermostat.

    Args:
        temperature: target temperature in Celsius
        mode: heating strategy to use
    """
    return {"temperature": temperature, "mode": mode}

agent = needle.Needle(tools=[set_thermostat])
agent.run("make it 21 and cool the room")

Example: You declare a weather function and ask what it is like in Lagos right now. Needle picks the weather call, runs your function and returns an answer built from the result. With the thermostat tool, asking for 21 degrees and cooling gives a call with those two values.

For newcomers and experts: A tool-calling model reads a request and decides which function to run and with what values. Needle 2 does this in a tiny package. A first small step is to install it and run the weather example. Describe your tools well, since the model reads those descriptions to decide what to call.

It is built on the authors' Simple Attention Network recipe and compressed to CQ2-bit with Cactus Quants. A grammar working at byte level, compiled from your schemas, limits what each token can be. A learned head gives each response a confidence score so you can set a threshold. A retrieval head shows only the top five tools each turn, and a 256-token sliding window keeps memory near 28 MB.

Declare tools as Python functions
Needle reads the request and your descriptions
Grammar constrains the call to your schema
Your function runs and the answer comes back

Good to know:

• Every response carries a calibrated confidence score, so you can act above a threshold and escalate below it.
• The card reports about 500 tokens per second on a Raspberry Pi 5.
• The card says it trades wins with other small models such as FunctionGemma 270M at 5 to 70 times smaller.
• License: apache-2.0.
• Cost to run: none named, and it needs no network.
• Data it reaches: the requests you give it and the tools you declare, with no downloads at run time.
• Not for: open-ended writing, since the card describes tool calling, device use and structured extraction.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: cactus-needle
• Downloads in the last 30 days: 1,736
• Likes: 341
• License: apache-2.0
• Fine-tunes: 10
• Tags: cactus-needle, needle, tool-calling, function-calling, on-device, edge, quantization, webassembly, text-generation

Links

• Model: https://huggingface.co/Cactus-Compute/needle2
• Files: https://huggingface.co/Cactus-Compute/needle2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Cactus-Compute/needle2

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