AWS releases Strands Decider 2B, a lightweight open-source

Amazon Web Services released the open-source Strands Decider 2B model, built on Qwen3.5-2B, as a high-speed, low-cost tool for choosing between pre-decided options. The model delivers a calibrated confidence score for its choice and is small enough to run locally, offering an alternative to full language models for discrete workflow steps. It does not generate text but focuses solely on decision-making and is part of AWS's broader Strands Agents program. This release occurred the same week OpenAI announced a similar offering.

Model origins trace back to engineer’s homebrew project inspired by TypeSafe’s J

Marc Brooker, an Amazon distinguished engineer, created Strands Decider 2B after seeing TypeSafe’s Jev. He initially built a homebrew version that briefly reached the top spot on the JevBench ranking for models of its size. AWS engineers later cleaned up that project for public release. Brooker said the need emerged from conversations with AWS customers whose agentic workflows do not always require the capability or cost of a fully featured LLM. He described the model as a perfect decider for a workflow step, asking 'what is the next thing for me to do here, based on where I am?' The model offers more reliability due to confidence scores and a closed domain of answers, with lower latency and potentially lower cost. Brooker noted the challenge is optimizing speedy decision-making without compromising intelligence, a balance between accuracy and preserving language understanding.

Strands Decider 2B achieves strong JevBench results and local deployability

Across the full JevBench benchmark, Strands Decider 2B ranks second among public models of roughly 2 billion parameters. AWS reports the model can make decisions in under 100 milliseconds on common hardware like the NVIDIA RTX 3090. The model's performance on specific benchmark tiers is detailed below.

The model is available for download on Hugging Face, with its training data and scripts hosted on GitHub. AWS fine-tuned the Qwen3.5-2B base model using LoRA (Low-Rank Adaptation).

Context: TypeSafe’s Jev and the rise of 'system one' models for focused decision

Both Strands Decider 2B and TypeSafe’s Jev, released in September 2026, belong to the emerging 'system one' model category designed for calibrated choices without text generation. TypeSafe named their model after economist William Stanley Jevons, hoping to invoke his theory that falling cost of computer intelligence increases demand. Dozens of similar models have emerged since Jev's debut. Diogo Almeida, CEO and founder of TypeSafe, downplayed current competition. "I get that people think it’s a gold rush, but they might be underestimating the difficulty of making the models actually smart," Almeida said. He added that he did not see real competition for TypeSafe emerging yet, characterizing current models as more like ML people implementing a cool architecture than teams dedicated to making intelligence useful. TypeSafe executives stated they are focusing on improving future models. This sector is seeing increased activity, with OpenAI recently previewing its Decisions API. The pointer-head approach scores predefined options, so open-ended answers still require traditional language models. A perfect score on JevBench's easy tier is promising but does not guarantee performance on harder, messier real-world decisions. Second place on the full benchmark leaves room for competitors. Brooker noted that smaller markets mean the cost to build something interesting is now in the hundreds or thousands of dollars. Strands Labs is the organization developing new tools and protocols for deploying AI agents.

Strands Decider 2B is available for download on Hugging Face with training data and scripts on GitHub.