Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash

Category: ai-ml

Tags: on-device-ml, tool-calling, tiny-models, structured-extraction, edge-ai

Score: 8.0/10 (Innovation: 8, Technical: 9, Documentation: 8, Utility: 7)

Needle 3 is an 8-29MB foundation model from Cactus Compute designed to run tool-calling, structured extraction and embedding tasks on tiny devices like mobiles, wearables and microcontrollers. It introduces a Laddered Simple Attention Network with Monarch Hadamard MLPs, engram n-gram memory and byte-level grammar-constrained decoding, allowing sub-30MB subnetworks to match models 10x their size on mobile tool-call benchmarks. The Python-first API, calibration head and per-platform prebuilt engines make on-device agentic workflows practical without cloud round-trips.

Target audience: backend devs, mobile/embedded developers, AI engineers, edge/IoT developers

Repository: https://cactuscompute.com/needle · Python · Apache-2.0 · 12056 stars

View on Hacker News