AWS Strands Labs Releases Strands Decider 2B: An Open Source Decision Model That Picks Options in About 115 ms

AWS Strands Labs has released Strands Decider 2B, an open source model built to make decisions, not draft paragraphs. Feed it a state plus typed questions and it returns a choice, a yes/no probability, or a score with a calibrated confidence. It runs locally on a CPU, a consumer GPU, or an Apple silicon Mac, and the weights are on Hugging Face under Apache-2.0.

This is not a chatbot wearing a fake mustache. The model does not generate text, and the team says it is worse than reasoning models on complex problems. It is aimed at places where the answer space is already boxed in: routing, tool selection, triage, guardrails, evals, and hybrid agents. For the bundled CLI and HTTP server, installation is just pip install strands-decider. But the server binds to 127.0.0.1 and ships with no authentication, so anyone thinking about production has to bring their own auth layer.

Under the hood, Strands starts from Qwen3.5-2B-Base, drops the language modeling head, and swaps in a small pointer head of about 1 million parameters. One forward pass decides the result, with no decoding loop. The request supplies the label set, so the option count is not capped. The released checkpoint is v19.

The numbers are decent and the calibration looks like the real selling point. On JevBench v1 public, v19 posts 0.723 accuracy, with a Brier score of 0.342 and an expected calibration error of 0.052. On an RTX 3090, median latency is 115 ms; on an M3 Pro, the warm median is 153 ms under 300 tokens. The repo also notes a catch: Mapika’s newer decider-2b v11 is ahead on the Strands harness, and Strands Decider was not on the newer v1.5.4 composite board at the time of writing.

The practical pitch is simple: if a model just needs to pick, don’t pay for it to talk. That is the part the industry keeps rediscovering after stuffing every problem into an LLM-shaped box.