Amazon released Strands Decider 2B, a model that chooses among supplied options or assigns a rating instead of writing an open-ended answer. Its code, weights and training materials are available for local use. The intended work includes selecting tools, routing requests and checking policies inside AI agents. The design scores alternatives in a single pass and is not intended for chat or complex reasoning.
Amazon’s Strands Labs team released Strands Decider 2B on 1 October as an open decision model. It chooses among options supplied by an application or assigns a rating, keeping the response inside predefined alternatives. The release includes code and downloadable weights for developers to run locally. Its intended setting is an AI agent’s workflow, where selecting a tool or routing a request can require a short, bounded decision.[1], [2]
The architecture starts with the Qwen3.5-2B model’s core and replaces its language-generation head with a pointer head. That component scores the available options in one forward pass, avoiding a sequence of generated words. The interface accepts yes-or-no questions, a selection from supplied choices, or a rating on an ordered scale. Amazon also released training data and scripts alongside the model.[1]
A worked example checks a weather-tool request when the user has not named a city. The decision model examines whether the proposed argument is grounded in the conversation and whether calling the tool is premature. Application code then asks the user which city they mean. Amazon identifies tool selection, routing and policy checks as uses, while excluding chat, document summarization, code generation and complex reasoning from this model’s intended role.[1]
