GoKawiilCactus releases Needle 3, an 8-29MB on-device model for automation tasks
Cactus launched Needle 3, a foundation model small enough to run as a single 8-29MB binary on phones, wearables, robots, smart home hubs and cars. Built on what the company calls a Simple Attention Network, it sacrifices general chat ability to specialize in three tasks: text embedding for local search, structured data extraction from messy text, and tool calling that maps spoken requests to app functions. The company claims it outperforms models ten times larger on mobile tool calls and matches models two to three times bigger on extraction.
GoKawiil's interpretation of the reporting above, not reported fact.
Running capable language models entirely on-device, without cloud round trips, addresses latency, privacy and offline reliability concerns that matter for robots, wearables and cars where connectivity can't be assumed. By trading broad conversational ability for narrow task performance at a fraction of the size, this approach could make embedded AI assistants far more practical for battery- and memory-constrained hardware than shipping scaled-down versions of general chat models.
- Needle 3 is a single 8-29MB binary designed for phones, wearables, robots, cars and microcontrollers.
- It focuses on three specialized tasks: text embedding, structured extraction, and tool calling, rather than general chat.
- Cactus claims it beats models 10x larger on tool calls and matches models 2-3x bigger on extraction tasks.
Raspberry Pi 5 (8GB) — If you're experimenting with tiny on-device models like Needle 3, a Raspberry Pi 5 is a great low-power platform to prototype offline tool-calling and smart home automation without cloud dependency. Its compact size and solid compute make it ideal for testing local inference workloads like embeddings and structured extraction.
See Raspberry Pi 5 (8GB) on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: cactuscompute.com, 2026-09-18
Published there as: “Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.