Amazon Open-Sources Strands Decider 2B for Low-Latency Local AI Agent Decisions
Amazon's Strands Agents team open-sourced Strands Decider 2B, a specialized 2-billion parameter model optimized for local decision-making in AI agent workflows. Built by swapping the text generation head of a Qwen3.5-2B base model with a 1-million parameter scoring pointer head, the model is available on GitHub and Hugging Face for CPU and GPU execution. Standard generative language models incur significant latency and compute overhead when an agent simply needs to choose a tool or select an action. By repurposing an LLM backbone into a fast decision ranker, Amazon enables responsive local agent systems that do not depend on external cloud APIs. The model fine-tunes the Qwen3.5-2B trunk using a rank-16 LoRA adapter combined with a tiny scoring head. On the JevBench benchmark, it outperformed all direct sub-2B competitors in accuracy and calibration, delivering a median decision latency of 113ms on standard local hardware and 153ms on an NVIDIA GeForce RTX 3090.
## BACKGROUND
Traditional LLMs generate text token-by-token, which introduces high latency for structural routing tasks in multi-step AI agents. Techniques like Low-Rank Adaptation (LoRA) enable parameter-efficient fine-tuning by modifying only a small set of added adapter weights rather than retraining the full base model.