AI
AWS releases Strands Decider, a local decision model to rival TypeSafe's Jev
Amazon's new open-source decision model lets developers run AI reasoning tasks locally, with full training data and scripts included.
The New Stack
Amazon Web Services has introduced Strands Decider 2B, an open-source decision model that competes with TypeSafe's Jev and other similar systems. Rather than relying on a hosted API, AWS is providing a downloadable model along with the underlying training data and scripts—a move that sets it apart from OpenAI's recently announced Decisions API, which runs as a cloud service.
Why decision models matter#
Decision models occupy a distinct niche in AI applications. They forgo the free-form text generation of large language models in exchange for selecting from predefined options or assigning numerical scores. This makes them particularly suited for tasks like routing user requests, choosing which tools an agent should invoke, evaluating whether outputs meet requirements, and validating proposed actions before execution. The constraint on possible answers prevents the model from inventing nonsensical options—a useful safeguard for production systems.
How Strands Decider works#
The architecture begins with Qwen 3.5-2B as the core language-understanding component. AWS stripped away the conventional text-generation head and replaced it with a scoring mechanism that evaluates supplied answer options instead. This scoring layer contains just over a million parameters and uses rank-16 LoRA (low-rank adaptation) to keep the model lightweight.
The model runs quickly—under 100 milliseconds on an Nvidia RTX 3090, and around 150 milliseconds for typical tasks on an M3 MacBook. Performance depends on task complexity, but the goal is consistent, dependable decisions paired with confidence scores that developers can use to determine whether to proceed with an action or seek additional input.
A practical example in action#
AWS's demo illustrates the model in a real agent workflow. When a user requests weather information without specifying a location, the agent makes a guess and prepares to call a weather tool. Before execution, Strands Decider validates whether the proposed arguments are grounded in the conversation and whether the agent possesses sufficient context to act. If not, the system can route the request back to the user for clarification, deny the call, request human confirmation, or provide feedback to the agent to try again.
