Someone fit a reasoning model into 14MB

A team called Cactus Compute just shipped Needle2, a 14MB language model that can run locally on devices most people carry around. Phones, wearables, smart home hubs, robots. The entire model is smaller than a high-resolution photo. The claim is that it handles agentic workflows. Tool calling, multi-step reasoning, the kind of tasks people usually throw at cloud APIs. Except this one runs offline, no server handshake, no latency spike when your internet drops. Fourteen megabytes. For context, GPT-2 small was 500MB. Llama 3.2 1B is around 1GB. This is two orders of magnitude smaller. The tradeoff is capability, obviously. You are not getting GPT-4 quality reasoning out of something that fits in the margin of your phone’s storage. But for specific tasks, constrained domains, it might be enough. The interesting bit is deployment. If you can ship a model this small, you can embed it in hardware that would never talk to the cloud. A doorbell that understands “is this person a delivery driver”. A wearable that logs meal descriptions without uploading your voice to a data centre. A robot vacuum that routes itself based on spoken instructions, no Wi-Fi required. Most edge AI right now is either keyword matching or a stripped-down classifier. This is claiming to be a general reasoning engine. I want to see the benchmark scores. I want to know what tasks it actually completes versus what it fumbles. The HN thread has people asking about quantisation methods, training data, whether it is a distilled model or trained from scratch. No answers yet. If the quality holds up, this is the kind of constraint that forces better design. You cannot brute-force your way out of 14MB. You have to pick your vocabulary carefully, prune hard, compress smarter. That discipline might produce more useful tools than throwing another hundred billion parameters at the problem.

---

Source: Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots