Amazon Web Services Inc.’s Strands Labs team has been playing around with an emerging class of lightweight artificial intelligence systems known as “decision models,” and it’s now making the fruits of that work available to the open-source community.
The company said today it’s releasing an open-source decision model called Strands Decider 2B, which is designed to make decisions rapidly by eliminating the need to generate text, which eats up vast amounts of tokens and increases response latency.
Strands Decider 2B has been optimized for local deployments and rapid experimentation, so developers can download and run it on their own laptops as well as public cloud environments. By giving developers access to a dedicated decision-making engine, AWS said, it hopes to accelerate the pace of agentic AI development more broadly.
Decision models, also known as “System 1 models,” are very different things from the large language models that everyone knows and either loves or hates. Whereas LLMs respond to inputs by generating text, code, images or video, decision models simply make decisions when they’re presented with a number of predefined choices, with no other outputs. Each decision they make is assigned a confidence score to help users judge how accurate the response is likely to be.
The omission of text generation means that decision models excel in terms of speed and low latency. The confidence scores assigned to each response are crucial, because the lack of text-generation capabilities means these models can’t explain their decisions.
Though decision models have been around for a while, they only started getting attention a couple of weeks ago following the emergence of a startup called TypeSafe AI Inc. and its subsequent release of Jev. It was designed to make fast, structured decisions that software and AI agents can use directly. Jev demonstrated how useful decision models can be, but AWS’s team believes it suffers from a number of structural shortcomings. For one thing, its parallel output structure means there are performance issues when it’s asked to perform intricate reasoning.
AWS decided to try to improve on the concept and Strands Decider 2B is its first real attempt at doing that. It’s built on top of a standard LLM, using Qwen3.5-2B’s base torso, but AWS swapped out the traditional “LLM head” for a tiny, customized “pointer head” that has a relatively minuscule 1 million parameters. It fine-tuned the Qwen3.5-2B torso using a rank-16 LoRA adapter to score hidden states of available choices directly against answer positions.
The cloud giant said it has repeatedly iterated on Strands Decider 2B, with today’s release only named v.20. The company said it deliberately chose the 2 billion-paremater scale because it believes this size provides just the right balance, making it small enough to run on a local machine with less than 150 milliseconds of latency, yet still powerful enough to make complex decisions. Strands Decider 2B certainly stacks up well in benchmarks, with AWS reporting that it achieved high-level performance in terms of accuracy and calibration on JevBench when compared with other open-source 2B models.
The model is available to download now via Hugging Face, and the full codebase, training scripts and examples can be found on its GitHub page.
AWS hopes the AI developer community will experiment with Strands Decider 2B and accelerate agentic tasks such as model routing, tool selection, context management, guardrail enforcement and policy classification. It also paves the way for the creation of “hybrid agents” that use decision models to make simple and repetitive choices and LLMs for complex reasoning challenges.