Amazon Web Services has released Strands Decider 2B, a small open source AI model designed to make fast decisions inside agent workflows without relying on a full large language model for every step. The model selects from predefined options, reports how confident it is in each choice, and is small enough to run locally.
AWS introduced Strands Decider 2B on October 1 through Strands Labs, its experimental organization for agent tools and protocols. The release comes as AI developers experiment with decision models that trade the open-ended text generation of LLMs for faster and more structured outputs.
A Smaller Model For Agent Decisions
Strands Decider is built from the base of Qwen3.5-2B, but its text-generation component has been replaced with a system that scores predefined choices. The model has about 1.9 billion parameters and can answer questions in tens to hundreds of milliseconds depending on the hardware.
AWS says the approach is suited to tasks such as model routing, tool selection, argument checking, triage, policy classification, guardrails, and evaluating AI outputs. Developers can also combine a decider with an LLM, leaving more complex reasoning to the larger model while routing simpler decisions through the smaller one.
Each decision includes a confidence score. AWS says this can help developers decide when an answer is reliable enough to continue automatically and when a workflow should ask a person or another model for confirmation.
AWS Builds On The Jev Approach
Amazon distinguished engineer Marc Brooker began experimenting with the architecture after seeing TypeSafe’s Jev, another model designed around making structured decisions instead of generating arbitrary text. His initial version briefly reached the top position among similarly sized models on the JevBench benchmark before AWS engineers prepared it for release through Strands Labs.
Brooker said conversations with AWS customers showed that agent workflows do not always require the capabilities or expense of a full LLM. A smaller decision model can provide lower latency and potentially lower costs for workflow steps where the possible answers are already known.
The model scored 167 out of 231 tasks on the public JevBench test at its reference configuration, according to the project documentation. AWS also reports median latency of 115 milliseconds on an Nvidia RTX 3090 and 153 milliseconds for shorter tasks on an Apple M3 Pro.
AWS Releases The Full Model Stack
AWS has made the Strands Decider code available under an open source license, alongside its model weights, training data, scripts, evaluation tools, and research history. The company says developers can run the model on CPUs, GPUs, and Apple silicon hardware or retrain it using consumer and data-center GPUs.
Decision models remain more limited than reasoning models because they cannot generate arbitrary text or work through complex problems in the same way. AWS says Strands Decider is therefore not intended for tasks such as coding, chatbots, or document summarization.
The release follows TypeSafe’s Jev and arrived during the same week that OpenAI introduced its own model built around a similar decision-focused approach. TypeSafe CEO Diogo Almeida told TechCrunch that producing the architecture itself is only part of the challenge, with future development depending on improving the models’ underlying intelligence while preserving their speed and calibration.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.