Amazon Releases Strands Decider 2B, an Open-Source Decision Model for Agentic Workflows

Amazon has released Strands Decider 2B, a small decision model optimized for fast experimentation and local development within the Strands Labs ecosystem. Unlike large language models that generate arbitrary text, decision models are designed to select from predefined options and assign numerical scores, making them faster and more reliable for specific agentic tasks like routing and guardrails.

Architecture and Performance

The model is built on a Qwen3.5-2B torso with the language modeling head replaced by a pointer head that scores hidden states at option positions. This architecture allows the model to run on local CPUs or GPUs, achieving a median latency of 115ms on an Nvidia RTX 3090 and 153ms on an M3 MacBook. On the JevBench public set, Strands Decider 2B ranks 3rd of 33 in the 2B class for accuracy and calibration, demonstrating competitive performance for its size.

Use Cases and Integration

Strands Decider 2B is intended for tasks such as model routing, tool selection, evaluations, and policy classification. It integrates with the Strands Harness SDK to enable hybrid agents where LLMs handle complex reasoning while the decision model manages rote decisions to reduce cost and latency. The model is released as open source on GitHub and Hugging Face, including all training data and scripts, and can be accessed via the strands-decider CLI.