AWS Strands Labs has released Strands Decider 2B, an open-source 1.9 billion parameter decision model distributed under an Apache-2.0 license. Unlike standard language models that generate free-form text, Strands Decider reads an input state alongside typed questions and outputs discrete option choices, probabilities, or calibrated scores without a decoding loop. The architecture replaces the traditional language modeling head with a 1-million parameter pointer head attached to a Qwen3.5-2B base model.

Designed to operate locally on CPUs, consumer GPUs, or Apple silicon, the model achieves latency of roughly 115 ms per decision. It is designed specifically for System One operational tasks, including model routing, tool selection, guardrails, and hybrid agent task delegation where complex reasoning LLMs are unnecessary.

Why it matters

  • Offers developers a fast, low-latency open-source option for non-text agent gating, tool selection, and request routing.

  • Reduces inference costs by replacing expensive full-scale LLM calls with lightweight local decision models.

Source: marktechpost.com