AWS Releases Strands Decider 2B, a Small Model for an AI Agent’s Routine Decisions

In the 16 months since Amazon Web Services (AWS) released Strands Agents, its open-source SDK for building AI agents, the project has steadily grown beyond that original toolkit. Since then, the company has added a labs program for experimenting with new approaches to agentic AI and released a harness for running agents, among other things. Its newest addition is a small model the Strands team built itself, designed to handle the quick, routine decisions agents make constantly so a larger and slower large language model doesn’t have to.

On Thursday, AWS released Strands Decider 2B. It’s a small decision model that puts the company in a corner of AI that has drawn a surge of interest among agent builders since TypeSafe AI introduced its Jev model last month. These decision models, sometimes called system one models, earned their appeal by doing very little. Where a traditional chatbot might answer a prompt with several paragraphs, a decision model is given a question with a short list of possible answers and picks one, or it puts a number on something, such as how positive a sentence sounds. The examples AWS used to illustrate the idea are intentionally mundane, like telling Zulu apart from English and Dutch, or recognizing that a request to turn on the lights has nothing to do with the coffee maker.

AWS said this narrower scope pays off: decision models are faster and more capable than an LLM of the same size, always land on one option, and can respond with very low latency.

Some might ask why AWS opted to build a decision model for Strands. Why not a large language model? By now, LLMs have become commoditized. Strands aims to facilitate agentic development, so why not focus on other areas besides building a general foundation model? The recently announced Strands harness offers a clue about what it’s about—AWS wants to build everything an agent needs except the brains. Strands Decider 2B is a small, specialized component built to work alongside the LLM inside the agent loop.

To create Strands Decider, AWS took a pre-trained Qwen3.5-2B (its LLM “torso”) and fine-tuned it using a lightweight technique called LoRA, which trains a small add-on layer instead of rewriting the entire model. The company removed the LM head, thereby taking away its ability to generate text. In its place, AWS attached what it calls a pointer head, a scoring component of just over 1 million parameters that weighs each answer option the torso produces.

The choice to keep Strands Decider at 2 billion parameters was deliberate. AWS said it did so to encourage experimentation. The small model makes it easier for developers to try out on a wider range of hardware devices. The company also believes 2 billion is a “sweet spot” since it’s just “small enough for experimentation, large enough to do meaningful work.”

So what can you do with a model that can’t generate text, yet can only answer multiple-choice questions? AWS shared that developers have been using small decision models for model routing, tool selection, evaluations, guardrails, memory, context management, and policy classification. There are potential implications for hybrid agents as well, as AWS is seeing AI agents use LLMs for hard decisions and decider models for “easier rote decisions.” This approach could reduce cost and latency. 

With a 2 billion-parameter model, Strands Decider is described as being “suitable” for running locally on a CPU or GPU and can return answers to meaningful questions “in tens of milliseconds.” In its evaluations, AWS found that its small decision model matched or outperformed other decision models, such as Typesafe’s Jev, LocalJev, and OpenJev. When measuring Strands Decider on how well it answered questions and how trustworthy its confidence scores are, AWS reported strong performance, placing second among public models of roughly the same size and first among public models with the full training recipe available. On latency, or how quickly decisions are made, AWS said Strands Decider 2B can execute local decisions in under 100 milliseconds “on widely available hardware.”

Strands Decider 2B is open source, and AWS has released the training data and scripts it used to build the model along with it. The code, a getting-started guide, and example projects are available on GitHub, while the latest versions of the model itself can be downloaded from Hugging Face.