What changed
On October 1, 2026 AWS's Strands team released Strands Decider 2B, an open decision model based on a Qwen3.5-2B backbone. Instead of generating text, it scores a supplied set of allowed answers and returns a bounded choice with confidence information. AWS also released the training data and scripts, making the implementation reproducible rather than API-only.
Why it matters
Agent systems repeatedly make small decisions—tool selection, routing, policy classification, memory choices and escalation—that do not require open-ended generation. A small local model can move those decisions off expensive frontier APIs, reduce latency and keep workflow state local. More importantly, AWS releasing the training recipe adds another independent implementation to a category that has quickly expanded from Jev into open models and OpenAI's Decisions API.
The model deliberately cannot write prose
Strands Decider starts from a Qwen3.5-2B backbone but removes the normal language-model output head. A roughly one-million-parameter pointer/scoring head instead evaluates the supplied answer options. The resulting model is built for closed-domain decisions rather than chat or text generation.
AWS is targeting the boring decisions inside agents
AWS describes uses including model routing, tool selection, evaluations, guardrails, memory and context management, and policy classification. Those are frequent agent-loop operations where a large generative model may add cost and latency without adding useful freedom.
The open recipe makes the category easier to test
AWS has released the model plus training data and scripts. The company reports local decisions in under 100 milliseconds on widely available hardware and competitive accuracy/calibration against other decision models. Those benchmark and latency results are vendor-produced and need independent replication.