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

AWS has launched Strands Decider 2B, its own take on the new crop of decision models that started with TypeSafe’s Jev a few weeks ago. OpenAI also jumped in on Tuesday with a Decisions API preview, but AWS is taking a different route: a downloadable model, plus the data and scripts used to train it.

The point of these models is not to write prose. They pick from options a developer supplies or return scores, which makes them useful for routing requests, choosing tools, judging outputs, and checking whether an action should go ahead. In AWS’s setup, Strands Decider is aimed at the sort of small but important choices agents make before they act.

Under the hood, AWS built the model on Qwen3.5-2B, which it calls the “torso,” then stripped out the text-generating head and replaced it with a pointer head that scores the available answers. That head has just over a million parameters, and the backbone uses a rank-16 LoRA adapter. The appeal is obvious: the model can’t invent an answer option that wasn’t offered. But AWS is also careful not to oversell that trick. Restricted choices do not make the model perfect, only more controlled.

The company’s example shows why that matters. A user asks for the weather without naming a place, the agent guesses a city and prepares a weather tool call, and Decider steps in to check whether the argument is actually grounded in the conversation. When it isn’t, the app pushes the agent back to ask for the missing city. The check runs through Strands’ intervention system, where developers can let a tool call proceed, block it, ask for human confirmation, or send feedback back to the agent.

AWS says the model was tuned for accuracy, calibration, and latency. On JevBench’s public set, it ranks second among public models at roughly 2 billion parameters, and first among public models that ship a full training recipe. The easy tier is clean sweep territory: AWS says Strands Decider 2B gets every question right there. It also reports sub-100 millisecond decisions on an Nvidia RTX 3090, and a median of around 150 milliseconds for small tasks on an M3 MacBook. The release is the second major iteration of the architecture, and AWS has left earlier versions in the repository so developers can see how it evolved.