Cactus Releases Needle 3: an 8-29MB On-Device Model Built for Tool Calls
airesearch12 · x · 2026-09-19
Cactus Compute released Needle 3, a foundation model for mobile, wearables, robots, smart home, automotive and microcontrollers— a single 8-29 MB binary built on their Simple Attention Network.
It trades away general chat entirely: every turn is a function call. With only 121M parameters it beats models 10x its size on mobile tool calls and matches 2-3x larger models on structured extraction.
Three capabilities: (1) tool calling—picks the right functions and fills arguments, returns an empty list rather than guessing; (2) structured extraction—declare a shape, get typed fields from messy text, with decode grammar guaranteeing parseability; (3) text embeddings for local search, matching and routing.
One set of weights works at every depth from 2 to 20 layers, described as an "intelligence ladder." Official demos include offline smart-home commands and nuanced robot vacuum instructions.
Related event: Cactus Releases Needle 3, an 8MB On-Device Foundation Model for Tool Calls(4 posts)→
More from Embodied
- How robot arms are controlled: an explainer for ML practitioners entering robot learning — Roger_M_Taylor · 2026-09-19
- Flying humanoid robot walks upside down on ceiling using real-time whole-body MPC — ___Mufasaa · 2026-09-19
- CMU 16-831 Intro to Robot Learning (Fall 2025) video lectures available — caglar_ee · 2026-09-19
- Building 3D-printable models with GPT-6 Astra + Blender MCP, and offloading geometry to Higgsfield to save tokens — Julianw20012 · 2026-09-19
- Disney's VertiGo Robot Uses Twin Propellers to Drive Straight Up Walls — lukas_m_ziegler · 2026-09-19
- Running Jev as a real-time robotics policy in MuJoCo, with a two-call decide-then-act split — facontidavide · 2026-09-19