Amazon Web Services has launched Strands Decider 2B, an open-source model designed for rapid decision-making within AI agents. Released on October 1, 2026, the model features 2 billion parameters and is based on the Qwen3.5-2B architecture, fine-tuned using the LoRA technique.
The model utilizes a specialized 'pointer head' component with approximately 1 million parameters. This mechanism evaluates predefined options and selects the best fit without generating full text, which reduces token usage and costs. It is capable of making decisions in under 100 milliseconds on hardware like the NVIDIA RTX 3090.
In testing on the JevBench benchmark, Strands Decider 2B achieved a perfect score on the easy tier. It currently ranks second among public models of similar size and first among those providing a full training recipe. The model is available on Hugging Face, with training scripts hosted on GitHub.
AWS positions this release as an open-source alternative to the Jev model from TypeSafe AI, which debuted in September 2026. Both tools fall into the 'system one' category, intended to handle routine tasks alongside larger language models. The project is part of the broader AWS Strands Agents initiative.
