GitHub radar
Cactus Needle 3: 8–29 MB Models for Offline Automation
Tiny models for device-side command recognition: fits on a phone, watch, or browser via WebAssembly. Needs task-specific fine-tuning to work well — available for $19 through their service.
Needle 3 is a family of small models (between 8–29MB) that can be deployed to perform tool invocation and structured extraction on edge devices such as mobile phones, wearables, home automation systems, and web browsers using WebAssembly. It’s a single weight model with all depths between 2–20 layers included, so you can choose the model size based on your available resources. Given a user request and a set of predefined tools, the model outputs a typed JSON response indicating which tool should be invoked along with its parameters. You can fine tune this model on your own data for $19 using the Cactus hosted service. As the founder mentions in this HN thread, if you spend time curating your tool list appropriately for your application domain, you will get decent performance out of this model.
Why a vibe-coder should care
You can deploy Needle 3 in your apps that require user input (e.g., smart home, voice assistants, local AIs) and process the requests locally (i.e., without making any network calls/API requests). Once fine tuned on your tool list, this model performs well on specific, repetitive tasks. Don’t expect it to work like a Swiss army knife though; it underperforms state-of-the-art models on general benchmarks by a large margin and can only support up to ~10 tools.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Go to https://cactuscompute.com/needle, read the fine-tuning documentation, and prepare 30 example commands for a smart home (lights, timers, music, temperature) in the required format. Show me the dataset before submitting.
Runs on any device — phone, watch, or browser via WebAssembly. No GPU or extra RAM required.
Open on GitHub▌ More finds