New Models AI Minute Newsroom 2026-08-12

45 million parameters, 14 megabytes, and it can still call tools — the model that runs on a microcontroller

Cactus released Needle 2 on 10 August: a 45-million-parameter language model built for agentic work — tool calling, operating a device, pulling structured data out of text — that ships as a 14 MB binary and runs a whole session in about 28 MB of RAM. The trick is quantisation applied from the very start of training rather than bolted on afterwards. Cactus trains weights, activations and the KV cache at 2 bits throughout, so the model that gets deployed is exactly the model that was trained, sidestepping the quality collapse that post-hoc 2-bit compression causes at this size. It decodes at over 500 tokens per second on a Raspberry Pi 5 with prefill above 800, runs at 400–1,500 on VR headsets and 300–700 on sub-$200 Android phones, and works on ESP32-S3 microcontrollers, Meta Quest 3S and Apple Vision Pro. Across five public tool-calling benchmarks it trades wins with models 5 to 70 times larger running at full precision: 63.7% on Mobile Actions against 69.1% for Liquid AI's LFM2.5 230M and 64.0% for FunctionGemma 270M. Weights are Apache 2.0 on Hugging Face and GitHub.

Why it mattersAlmost every voice assistant in a doorbell, a car or a pair of headphones today is a thin client for someone else's data centre — which is why they stop working without a signal, why they cost the manufacturer money per interaction, and why what you say passes through a server. A model this size changes the arithmetic of that decision, because 14 MB fits in flash memory on hardware that costs a few dollars. Needle 2 is not clever and cannot hold a conversation; it is built to understand an instruction and press the right button, which is most of what these devices actually need. Apache 2.0 means anyone can ship it, and the interesting consequence is not smarter gadgets but gadgets that keep working, and keep quiet, when the internet is gone.

#AI Agents

✓ Verified · 3 sources

Read in the app — free, in 9 languages

45 million parameters, 14 megabytes, and it can still call tools — the model that runs on a microcontroller

Related stories