Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots
Points and comments are a snapshot, not live.
Cactus Compute releases Needle 2, a 45M-parameter LLM in a 14MB binary for tool calling on sub-$200 devices.
Needle 2 is an open-source, 14MB, 45M-parameter agentic LLM for tool calling and structured extraction on tiny devices. It runs on a Raspberry Pi 5 at 500+ tokens/sec decode, fitting in 28MB RAM through CQ2-bit quantization and a co-designed inference engine. Benchmarks show it trading wins with models 5x to 70x larger on mobile action and tool call tests. Targets sub-$200 devices like phones, wearables, and microcontrollers, with an Apache 2.0 license.
What commenters are saying
Commenters were impressed by the model's tiny size and specialized focus, but several tested the web demo and found false positives, especially a recurring 'lock_door' default. The Cactus team explained that the confidence score is intended to threshold such responses. A key discussion emerged around pairing Needle 2 with speech-to-text (e.g., Whisper) for screenless devices, with Home Assistant recommended as the best entry point.