A decision model answers narrow questions instead of writing paragraphs, and AWS’s Strands Labs has released one small enough to sit on a laptop. Strands Decider 2B takes a state plus typed questions and comes back with a pick from the allowed options, a yes-or-no probability, or a score that carries a confidence figure. Prose never enters the picture.
The 1.9B-parameter model runs on a CPU, a consumer GPU, or an Apple silicon Mac. Weights sit on Hugging Face under Apache 2.0, and installing the package yields both a command-line tool and an HTTP server, though that server has no authentication and binds to localhost only. No hosted inference provider offers it yet.
The build swaps out language modeling. Starting from Qwen3.5-2B-Base, the team drops the language head and bolts on a pointer head of roughly one million parameters that compares hidden states across options in a single forward pass.
The practical draw is calibration. Sweep the unseen short-classification set and a pattern shows up: whenever the model reported 0.9 confidence or better, it was right roughly 95 percent of the time. Under that bar, the guidance is to verify the answer or bump the question up a level. Third-party testing put the version-19 checkpoint third among 33 models in the 2B class on the September 25 JevBench board; set aside three entrants that run slightly larger, and it tops 30. Coding, chat, and summarization are explicitly out of scope; routing, tool selection, and guardrails are in.
