For an AI agent, the next move is often a multiple-choice question, not an invitation to write an essay. Amazon Web Services has released Strands Decider 2B, an open-source model designed to choose among set options and report how confident it is, rather than generate text.
A model with a narrower job
Strands Decider 2B is built on Qwen3.5-2B, but its output is a decision from a defined set of answers. Amazon presents it as a fast, low-cost option that is small enough to run locally.
The intended use is a step inside an agent workflow. An agent might need to decide what to do next based on its current state, without asking a full-featured language model to handle every turn. A bounded set of choices and a confidence score can make that step more structured, and potentially lower its latency and cost.
Those are the model’s aims, not a published performance guarantee. The report gives no benchmark results for Strands Decider 2B beyond its early showing on Jevbench.
From a personal experiment to an AWS release
Amazon distinguished engineer Marc Brooker began the project after encountering Jev, a decision model from TypeSafe. His own version briefly reached the top of Jevbench among models of its size. Amazon engineers then prepared it for release through Strands Labs, which develops tools and protocols for deploying AI agents.
The launch came in the same week that OpenAI announced a similar offering. Strands Decider 2B also joins a growing group of models inspired by Jev, with dozens of related efforts appearing since TypeSafe introduced the idea.
The trade-off behind the small model
A decision model has to be quick and accurate, but narrowing its job too aggressively can weaken the qualities that make it useful across different tasks. Brooker described the challenge as preserving accuracy and well-calibrated confidence without losing language ability or general knowledge.
That tension matters because a model that reliably chooses from a fixed menu is useful only when the menu and the situation make sense to it. The confidence score is part of the appeal, but it does not remove the need for the model to understand what it is being asked to decide.
Brooker also argued that smaller markets do not necessarily belong to the biggest AI labs. He put the cost of building something interesting in this area at hundreds or thousands of dollars, though the report gives no further breakdown of that estimate.
TypeSafe founder and CEO Diogo Almeida said he was not yet seeing serious competition for his company. In his view, many new projects reflect interest in the architecture, while the harder work is making the resulting models genuinely useful.
That is the test for Strands Decider 2B too. A smaller model can make a workflow cheaper or quicker, but only if its narrow answer is still the right one.