Cactus runtime

Cactus runtime

Provides prebuilt runtime binaries that run text, speech and vision models locally on phones, laptops, servers and browsers. For developers who ship on-device AI in apps.

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: Cactus runtime is a set of prebuilt binaries from Cactus Compute. One engine runs text, speech and vision models on the device itself. This repository holds the engine for 17 targets, and the model weights for Needle 3, Whistle and Mirage live in separate repositories.

Who it suits: Developers who want to put a model inside an app or a device, and who need one engine for many platforms.

Industry | Who uses it | How they use it | Example
Mobile apps | iOS and Android developers | Run a model on the phone itself | A speech clip transcribed on a phone
Desktop software | Desktop app developers | Embed the engine in an application | A native runner on Windows or macOS
Embedded and IoT | Hardware teams | Run models on ARM, RISC-V and MIPS devices | A model on a Linux ARMv7 board
Web apps | Front-end developers | Run the engine in a browser or Node.js | A web page using the WebAssembly build

If you only want to chat with a model, or need the weights themselves, you can skip it, since they sit in other repositories.

Where to use it:

• Python, through the cactus-compute package
• macOS, Windows, Linux, Android, iOS, tvOS and watchOS
• A browser or Node.js, and WASI Preview 2
• C and C++ applications, through the cactus.h header

Use case | Who | What they get back
On-device inference | App developers | A native runner that loads .cact weights
Embedding in C or C++ | Native developers | A static library and a header to link against
Python scripting | Python developers | The engine and weights fetched on first use

How to set it up:

• Python and pip for the Python route
• A target folder that matches your platform, for the native route
• Model weights from their own repositories, or fetched for you on first use in Python

1. Install the Python package.
2. Use import cactus in your code.
3. For a native runner, download one for your platform together with a model.

pip install -U cactus-compute

The files can also be fetched directly from Hugging Face.

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

How to use it: With Python, the first call fetches an engine that matches your machine, plus the model weights. For a native runner, download it with a model, then run it on an input. The card's example downloads a runner for Apple Silicon with the Whistle model and runs it on an audio clip.

cactus download macos-arm64 --model whistle --out ./runtime
./runtime/macos-arm64/cactus --model ./runtime/macos-arm64/whistle.cact --audio clip.wav

Put needle3 or mirage after --model to pick one of the other two models. Leave out --model to download the binaries alone.

Tool | What it does
cactus download | Fetches a runner for a platform, with an optional model
cactus runner | Runs a .cact model file on your input
cactus-compute package | Gives Python access to the engine

Example: You download the macOS Apple Silicon runner with the Whistle model and run it on a short audio file called clip.wav. The runner loads the model file and processes the audio on your machine.

For newcomers and experts: If you are new to this, install the Python package and try the model through Python, which fetches what it needs. Native runners are for when you want to ship an app.

The release has 17 target folders. Native runners load .cact weights at runtime, and static libraries ship with a C header. The Python wheels cover macOS, Linux with glibc or musl, and Windows. A WASI component is also offered, and release files carry SHA-256 checksums in a manifest.

Pick the target for your platform
Download the runner with a model
The runner loads the .cact weights
It processes your text, audio or image

Good to know:

• Weights: this repository holds the engine only, and each model's weights are in a separate repository
• Platforms: iOS, tvOS and watchOS targets give a static library and a header, not a runner
• License: apache-2.0
• Cost to run: no API key or paid service is named
• Data it reaches: the Python package downloads the engine and weights on first use, and runners read the files you give them
• Not for: use as a model by itself, since you need the weights from Needle 3, Whistle or Mirage

On Hugging Face, 13 November 2026

• Task: text generation
• Library: cactus-compute
• Downloads in the last 30 days: 682
• Likes: 19
• License: apache-2.0
• Tags: cactus-compute, text-generation, automatic-speech-recognition, image-text-to-text

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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