Cactus Compute

Cactus Compute builds very small foundation models that run locally on devices, led by Needle, now in its third version. Needle is a single 8 to 29 MB binary that trades general chat ability for reliable automation: given the functions an app exposes, it chooses the right tool calls and fills their arguments from what the user said, extracts typed fields from messy text against a declared schema, and returns text embeddings for local search and matching. It is aimed at developers building voice and text control into hardware and apps where cloud calls are slow, costly or unavailable: phones, wearables, AR glasses, smart home devices, robots, cars, computers, game consoles, TVs and microcontrollers. The model uses a laddered Simple Attention Network in which every depth from 2 to 20 layers is a usable model, is quantized to two bits and, according to the vendor, decodes at up to 4,000 tokens per second on a Raspberry Pi 5. Output is structured JSON constrained by a decode grammar so it always parses. A Python package fetches the inference engine from Hugging Face, and developers can define tools with decorated functions and route requests with regular expression triggers. Fine tuning pricing is published in US dollars. Local fine tuning with the open source package is free, with no account required. Hosted Starter costs $19 once for 30 days of access, three fine tuning runs, 3,000 generated examples and 10 GB of storage, and is available once per account. Hosted Pro costs $99 per month for ten runs and 10,000 generated examples per period. Custom allowances for higher volumes are quote based. Cactus is backed by Y Combinator, and the Needle weights and code are published on Hugging Face and GitHub.