AWS-affiliated Strands Labs releases open Decider 2B model
The small, locally runnable model scores supplied choices for agent workflows instead of generating open-ended text.
A model for the next bounded decision
Strands Labs, an experimental group linked to AWS’s Strands agent tooling, released Strands Decider 2B on October 1. It is a two-billion-parameter model built to score options supplied by an application rather than write arbitrary text. A developer can ask it to select a support team, classify a request or assess a proposed tool call before an agent acts. The team says its model weights, source code, training data and scripts are public, and that it is small enough for local use. TechCrunch separately reported the release and interviewed Strands engineer Marc Brooker about use in agent workflows.
The Strands team says it adapted a Qwen3.5-2B base by replacing the language-model output head with a choice-scoring head. It reports benchmark results for accuracy and confidence calibration, plus local latency measurements on particular hardware. Those figures are the team’s own tests; they do not establish accuracy, latency or dependable confidence scores for every real customer task. A narrow option set also means a developer must define the possible answers carefully.
Where it fits
The official announcement names routing, guardrails and policy classification among possible uses. It also states a clear limit: a model that must choose from given answers is not suited to chat, coding or document summarization, and it is weaker than reasoning models on complex problems. Its value is in cheap, structured decisions inside a larger workflow, not in replacing a general-purpose assistant.