Model
Mirage
A 22 MB vision model with object detection, OCR, segmentation, decision and embedding.
Try in-browser790/moHugging Face downloads last month
Backed by
Sub-30MB Foundation Models for Mobile, Wearables, AR glasses, Smart homes, Robotics, Automotive, Laptops, Game consoles, TVs, Microcontrollers
Agents, voice and vision running locally on consumer hardware.
Latest releases
- Model
Whistle
A 16.9 MB open speech recognition model that supports seven languages and runs on CPU only.
Try in-browser75.2k/moHugging Face downloads last month - Model
Needle 3
Our latest Automation Foundation Model. A single 8–29 MB binary that beats models 10× its size on mobile tool calls.
Try in-browser118k/moHugging Face downloads last month - Company
Pebble partnership
Cactus now powers on-device intelligence for Pebble watches and rings.
- Research
Hybrid routing
Post-trained models that know when they're wrong and request help from the cloud.
Read more
Demo
Needle, Whistle and Mirage
Try our models directly in your browser: tool calling with Needle 3, speech recognition with Whistle and vision with Mirage.
SandboxRuns Locally In Your Browser
Capabilities
Big models, small hardware.
Our research that enables models to run on everyday devices, and fall back to the cloud only when needed.
Sub-2-bit quantization
We compress foundation models below two bits per weight with almost no quality loss, so large models fit in the memory of everyday devices.
- Weights
- < 2 bits
- Quality
- Near-lossless
- Models
- Ours + third-party
Microprocessor-tier models
Our smallest models run on budget phones, wearables, and Raspberry Pis. No GPU, no server, and no data leaving the device.
- Minimum hardware
- CPU-only
- Network
- Offline + Hybrid
Hybrid routing
Our models report how confident they are with every answer. That score decides whether a request stays on the device or goes to the cloud.
Trusted by the industry
“The Pebble Index Ring has no screen. So when you speak to it, the action just has to happen, every time, with or without internet connection. We run Cactus Needle locally in the app, instead of relying on the cloud. The model's footprint is tiny and the performance never lets us down.”
1 / 2