AWS Debuts Strands Decider 2B for Agentic Workflows

Amazon Web Services Inc.’s Strands Labs team has introduced Strands Decider 2B, an open-source lightweight decision model built to accelerate agentic workflows. By removing the need to generate text, the new model aims to cut down token consumption and response latency, giving developers a dedicated engine for rapid choices.

Unlike traditional large language models that respond by generating text, code, or images, decision models—sometimes referred to as System 1 models—simply select from predefined choices. Each output comes with a confidence score to help users gauge accuracy, though the absence of text generation means the model cannot explicitly explain its reasoning. Attention around this category of models grew recently following the launch of startup TypeSafe AI Inc. and its Jev system. AWS noted that while Jev demonstrated the utility of decision models, it suffered from structural shortcomings, prompting the cloud giant to build its own iteration.

Architecture and Performance

Strands Decider 2B is constructed on top of a standard LLM, utilizing the Qwen3.5-2B base torso while replacing the traditional LLM head with a specialized pointer head containing roughly 1 million parameters. AWS fine-tuned the base torso using a rank-16 LoRA adapter to directly score hidden states of available choices against answer positions. Reaching version v.20 through multiple iterations, AWS selected the 2 billion-parameter scale to hit a balance between local machine compatibility—enabling performance with under 150 milliseconds of latency—and decision-making capability. Benchmark results on JevBench showed that the model achieved strong accuracy and calibration compared to other open-source 2B models.