Amazon Open-Sources Strands Decider 2B Decision Model That Runs on Local CPUs and GPUs

Amazon's Strands Agents team has launched the Strands Decider 2B decision model. It is now open-sourced on GitHub, with weights available on Hugging Face, and can run locally on CPU or GPU.

The model uses a pre-trained Qwen3.5-2B as its backbone, replacing the original text-generation language model head with a scoring pointer head. The new head is small, with just over one million total parameters, while the backbone is fine-tuned through a rank-16 LoRA adapter.

On the public JevBench dataset, Strands Decider 2B shows good accuracy and calibration, ranking third among 2B-level models and outperforming all competitors strictly under 2B. When running locally on common hardware, it delivers a median decision latency of 113ms. On an NVIDIA GeForce RTX 3090, the median latency for small decision tasks is 153ms.