Amazon releases open-source Strands Decider 2B for agent decisions

Amazon has released Strands Decider 2B, an open-source decision model designed to sit inside an agent workflow and score predefined options in a single forward pass. The roughly 2-billion-parameter model is fine-tuned from Alibaba's Qwen3.5-2B base.

According to AWS, the model uses an architecture called "Hobson," which removes the next-word prediction component and replaces it with a pointer mechanism that scores supplied answer options. Developers applied a rank-16 LoRA update and trained the new scoring component, adding just over one million parameters. The model is available for free on Hugging Face under an Apache 2.0 license, along with training code, data, and scripts.

AWS says the model can make local decisions in tens of milliseconds on short tasks and under 100 milliseconds on some common hardware and inputs. On the public portion of JevBench, version 19 reached roughly 72% accuracy with a Brier score of 0.35. AWS claims it ranks second among public models of similar size and is the first to publish a complete training recipe.

Strands Decider 2B is part of Strands Labs, AWS's experimental agent-development project. For now, AWS offers only the open-source model, with no hosted API or per-call fee.