News from the IT community on October 3rd: The Amazon (Amazon) Strands Agents team announced the launch of a Strands Decider 2B decision-making model on the 1st of this month, local time. The model has been open-sourced on GitHub, and its weights can be found on Hugging Face; it can also be run locally on CPU / GPU.

Strands Decider 2B is based on a pre-trained Qwen3.5-2B "trunk", and replaces the original language model's "head" which has text generation capabilities with a "head" that includes a scoring function. The new "head" is smaller in scale, with a total of just over one million parameters; the "trunk" is then fine-tuned through an rank-16 LoRA adapter.

Strands Decider 2B performs well in terms of accuracy and calibration on the JevBench public dataset. It ranks 3rd among 2B-level models, outperforming all competitors that strictly do not exceed 2B.

When running this model locally on commonly available hardware in the market, Strands Decider achieved a median decision-making latency of 113ms. On NVIDIA GeForce RTX 3090, its median latency for performing small-scale decision-making tasks was 153ms.