AWS launches a local answer to TypeSafe’s Jev decision model

AWS has introduced Strands Decider 2B, its response to decision models like Jev, Kev, imajev, Laya, and others. TypeSafe's Jev initiated this wave of decision models, prompting major AI vendors to release their own versions. OpenAI, for instance, launched its Decisions API as a preview, but AWS is offering a downloadable model, data, and training scripts.

Decider models prioritize selecting from developer-supplied options or providing numerical scores over free-form text generation. This makes them suitable for routing natural language requests, selecting tools, evaluating outputs, and validating proposed actions, while allowing generative models to handle conversation and complex tasks.

Strands Decider uses Qwen3.5-2B as its language-understanding foundation, known as the "core." The team replaced the text-generating head with a scoring head that assesses provided answer choices. This head comprises just over a million parameters and employs a rank-16 LoRA adapter. Limiting the answer space stops the model from generating non-supplied options, but it doesn't guarantee perfect accuracy.

In AWS's demonstration using the open-source Strands agent framework, a user asks for the weather without specifying a location. The agent guesses a city and suggests calling a weather tool. Before the tool executes, Decider verifies if the argument values are grounded in the conversation and if the agent possesses sufficient information.

If needed, the agent returns to ask for the correct city. The verification process then proceeds through Strands' intervention system, allowing developers to choose between proceeding with the tool call, denying it, requesting human confirmation, or providing feedback to the agent. In this case, Decider runs locally while the agent calls its generative model via Amazon Bedrock.

AWS is also developing integration libraries for decision-models. Building on an open Qwen model, like Kev, Strands Decider illustrates how much of this experimentation now relies on open weights. Kev supports local deployment and fine-tuning, and AWS's release provides developers with the recipe and scripts to adapt it to their specific tasks.

Like Kev, Strands Decider uses an open Qwen model, highlighting the importance of open weights in this field. AWS emphasizes balancing accuracy, calibration, and latency. Calibration refers to how closely the model's confidence scores align with its actual correctness. On JevBench's public set, AWS reports Strands Decider ranks second among models with roughly 2 billion parameters, and first among those with a complete training recipe available.

AWS notes that Strands Decider 2B correctly answers every question in JevBench's simple tier. This type of routine agent decision is what the model was designed for. The company also mentions that the first iteration of the architecture underperformed significantly. Moreover, AWS retains all previous iterations in the repository, enabling developers to follow the model's evolution.

Strands Decider originated at Strands Labs, AWS's incubator for experimental agentic AI approaches, which launched recently. It also follows the company's recent release of Strands Harness, which bundles the tools and supporting infrastructure required to run longer-lived agents.

It remains to be seen if AWS will also provide a hosted version of this model, including future versions, in its cloud. While hybrid scenarios are ideal for testing and running locally, a hosted version would be beneficial for production applications built on this model.