Amazon's Strands Agents team released the Strands Decider 2B decision-making model on October 1. The model is available as open source on GitHub, with weights hosted on Hugging Face. It is designed to run locally on standard CPU and GPU hardware.
The architecture utilizes a pre-trained Qwen3.5-2B base. The original text generation head was replaced with a scoring-focused pointer head containing just over one million parameters. The base model was fine-tuned using a rank-16 LoRA adapter.
In performance testing on the JevBench dataset, the model ranks third among all models with 2B parameters or fewer. It demonstrated high accuracy and calibration levels compared to its competitors in the same size category.
The model achieves a median decision latency of 113ms on standard hardware. When running small decision tasks on an NVIDIA GeForce RTX 3090, the median latency is 153ms.
