Amazon Web Services (AWS) has introduced an open-source model called Strands Decider 2B for the decision-making steps of AI agents. Inspired by TypeSafe’s Jev model, the model selects among predefined options instead of generating text and adds a confidence score to its decisions. Small enough to run on local devices, the model is based on Qwen3.5-2B. The model is positioned for limited tasks, such as an agent selecting the next action based on its current state.
The experimental project developed by AWS engineer Marc Brooker briefly rose to first place in the Jevbench ranking in its size category and was then released under the Strands Labs umbrella. Discussions with AWS customers showed that using large models for simple steps increases latency and cost.
Following the introduction of Jev, dozens of similar models have been developed; OpenAI also announced the Decisions API this week. TypeSafe CEO Diogo Almeida, meanwhile, says the real challenge is making the models useful and intelligent, noting that they do not see serious competition at this stage. It is assessed that small decision models could provide advantages in terms of cost and latency in narrow tasks such as routing, classification, and selecting the next action.
Why it matters
As AI agents become more widespread, the cost and slowness of calling large language models at every step are becoming more visible; this picture, which emerged in AWS’s customer conversations, shows that the idea of moving decision-making to a separate and inexpensive layer was born out of a practical need. OpenAI announcing the Decisions API in the same week and the emergence of dozens of similar models after Jev suggest that narrow-task decision models are not a passing experiment but are turning into a new competitive arena where major players are also positioning themselves. This development concerns most of all the teams building agent-based software and those who want to run latency-sensitive applications on-device, because the model being open source and able to run locally opens the way for experimentation without cloud dependency. The remaining open question, meanwhile, hinges on the point the TypeSafe CEO highlighted: although getting the model to choose among options appears technically solved, how useful and intelligent these models will behave in real products has yet to be proven.
Background
Amazon is not a new name in the FikirPilot archive: we have published 12 stories mentioning this name in the last 90 days; the most recent is dated October 2, 2026.
Term: agent
An AI agent is software that, rather than producing a single response, calls tools and carries out multi-step work to achieve a goal.