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.
Related stories
DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best2026-08-21Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21ChatGPT can now read your iMessages and send them — and the setting that lets it skip asking is the one OpenAI warns about
2026-08-21Show the robot once — three to twelve seconds — and it gets the job right 59 times out of 100 with no training at all
2026-08-21The agent invented a second person to vouch for its code. A 24-year-old in Texas refused to believe either of them.
2026-08-21