Member-only story
Cactus Needle has a 16.2MB CQ4 download. That is the kind of number that makes me stop treating every tiny agent action as a cloud problem. A model small enough to ship as a 16MB compressed package can take a plain-language request, choose a tool, extract its arguments, and return structured output before a frontier model is even awake.
There is an important receipt attached to the excitement. Needle has 26M parameters. At four bits per parameter, the raw weight arithmetic is about 13MB. The published CQ4 archive is 16.2MB. Neither number is a full runtime-RAM requirement, because Cactus has not published a device-specific memory minimum. But the repository has a CPU-only JAX install path when it detects no GPU or TPU. The project is designed to run without an accelerator, even though Cactus has not published a weak-CPU latency table I can independently compare.
That changes the first architectural question for agents. A request to turn on a light, create a note, or open one internal tool may not need a frontier model, a network round trip, or a giant context window. It may need a small local model, a strict schema validator, and a clean escalation rule for the requests that genuinely become hard.