AWS Releases Strands Decider 2B, an Open-Source Model for AI Workflow Decisions

Amazon Web Services has released Strands Decider 2B, an open-source decision model designed to help developers automate choices within AI workflows.

Rather than generate an open-ended response, the model selects from predefined options and returns a measure of confidence in its decision. AWS positions it as a high-speed, low-cost tool for tasks that do not always need the capabilities of a larger language model.

The release arrives amid growing interest in decision-focused AI, following TypeSafe’s introduction of Jev and a similar offering announced by OpenAI this week.

A Smaller Model for Structured Decisions

Strands Decider 2B is built on Qwen3.5-2B, using the underlying language model to interpret information before producing a structured choice.

Its output is constrained to a defined set of answers. That makes it suitable for workflow steps where a system needs to determine what to do next rather than compose a paragraph.

AWS distinguished engineer Marc Brooker said customer discussions highlighted that need. Agentic workflows often contain decisions that do not justify invoking a fully featured language model every time.

A smaller decision model could handle those steps with lower latency and potentially lower costs, while confidence scores provide additional information about the strength of each choice.

The model is available now and small enough to run locally.

A Personal Experiment Became an AWS Release

Brooker began the project after encountering TypeSafe’s Jev and attempting to build his own version.

His initial implementation briefly reached the top of the Jevbench rankings for models in its size category. Amazon engineers subsequently refined the project and released it through Strands Labs, which develops tools and protocols for deploying AI agents.

The result turns an experimental implementation into an offering developers can examine and use.

Its open-source availability also gives teams room to evaluate the model within their own systems, rather than rely solely on a hosted service.

Confidence Scores Support More Controlled Workflows

A central feature is the model’s ability to pair its selection with a confidence score.

Brooker argues that this combination can make workflow decisions more reliable: the system operates within a closed set of answers and provides a signal about uncertainty.

That signal could help developers distinguish straightforward decisions from cases requiring additional review or more capable processing.

However, the usefulness of confidence scores depends on calibration—whether the reported confidence accurately reflects how often the model is correct.

A constrained answer format can simplify integration, but it does not eliminate the possibility of choosing the wrong option.

Accuracy and General Understanding Remain Challenges

Brooker said the engineering challenge involves improving decision accuracy and calibration without weakening the model’s broader understanding.

Developers still need it to comprehend different languages and draw on general knowledge. Optimizing solely for speed or narrow benchmarks could reduce those capabilities.

Dozens of similar models have emerged since TypeSafe introduced Jev, demonstrating interest in the approach while raising questions about how products will distinguish themselves.

Brooker does not assume frontier AI companies will dominate. He said smaller-scale projects can be built for hundreds or thousands of dollars, allowing more teams to experiment.

TypeSafe Says Architecture Alone Is Not Enough

TypeSafe CEO Diogo Almeida said his company is concentrating on improving future models rather than reacting to the growing field.

He argued that new entrants may underestimate the difficulty of making decision models genuinely capable, describing many current efforts as architectural experiments rather than sustained work on useful intelligence.

The competing releases show growing demand for AI designed around specific automation needs. Whether Strands Decider becomes a dependable workflow component will depend on its accuracy, calibration and performance in developers’ own applications.