Amazon Web Services on Thursday introduced Strands Decider 2B, a decision model that answers every question correctly in JevBench's easy tier and delivers results in under 100 milliseconds on an Nvidia RTX 3090 graphics card. The release joins a growing field of decision models from major AI vendors, but unlike OpenAI's hosted Decisions API launched Tuesday, AWS is providing a downloadable model along with the complete training data and scripts. Decision models swap open-ended text generation for choosing among developer-provided options or producing numerical scores, making them useful for directing natural language requests, picking tools, assessing outputs, and verifying proposed actions before execution.
Strands Decider uses Qwen3.5-2B as its language-understanding foundation, which AWS labels the "torso," then strips away the language-model component that produces text and substitutes a pointer component that evaluates supplied answer choices. That pointer component contains just over one million parameters, and the backbone employs a rank-16 LoRA adapter. On JevBench's public set, Strands Decider ranks second among public models with roughly 2 billion parameters and first among public models offering a complete training recipe. Response times climb as tasks grow, with the median for small tasks on an M3 MacBook sitting around 150 milliseconds, according to the company. AWS left every earlier iteration in the repository so developers can follow how the model changed over time, noting that an earlier head design performed significantly worse than the current architecture.
The company says it concentrated on balancing accuracy, calibration, and latency, with calibration referring to how tightly the model's confidence scores correspond to how often it actually produces correct answers. AWS developed the model to prevent agents from inventing options that weren't provided, though that constraint doesn't guarantee correct answers every time. In the company's demonstration built with its open-source Strands agent framework, a user requests weather information without specifying a location, the agent assumes a city and proposes calling a weather tool, and before that tool executes, Decider verifies whether the argument values are grounded in the conversation and whether the agent has sufficient information to continue. The application then directs the agent to ask which city the user intended.
Decision models trade the flexibility of free-form generation for speed and verifiable confidence scores that developers can use to gate agent actions, a particularly valuable capability as autonomous systems handle tasks with real-world consequences. That check operates through Strands' intervention system, which allows developers to choose whether to proceed with a tool call, reject it, request human confirmation, or send feedback to the agent, while Decider runs locally and the agent calls its generative model through Amazon Bedrock. Like competitor models including Kev, Strands Decider builds on an open Qwen model, demonstrating how much current experimentation relies on open weights that support local deployment and fine-tuning. Strands Decider emerged from Strands Labs, AWS's unit for experimental approaches to agentic AI that launched earlier this year, and follows the company's recent release of Strands Harness, which bundles the tools and supporting infrastructure needed to operate longer-lived agents. One uncertainty is whether AWS will also provide a hosted version of this model or a future version in its cloud, since hybrid scenarios work well for experiments and local development but production applications typically require hosted versions. The decision to publish not just model weights but the full training pipeline signals a bet that transparency and customizability will drive adoption faster than keeping the recipe proprietary. For organizations building multi-step AI workflows, the real competitive edge may lie less in raw model performance than in how cleanly these lighter decision layers integrate with existing infrastructure and governance frameworks.