Amazon Web Services launched Strands Decider 2B on October 1, an open-source model engineered to handle discrete decision-making in AI agent workflows with minimal computational overhead.

The core innovation is its pointer head—a lightweight 1-million-parameter component that scores predefined options rather than generating text token-by-token. This architectural constraint delivers two concrete economic advantages: it cuts token consumption and slashes operational costs, while hitting sub-100-millisecond latency on commodity hardware like the NVIDIA RTX 3090.

AWS tested Strands Decider 2B on JevBench, a decision-model benchmark. The model ranked second among public 2-billion-parameter models on the full benchmark and first among models released with full training recipes and reproducible code. Both the model and training scripts are available on Hugging Face and GitHub.

Under the hood, AWS fine-tuned Alibaba's Qwen3.5-2B base model using LoRA, a parameter-efficient technique that adjusts only a small weight set rather than retraining the entire model.

The release directly challenges TypeSafe AI's Jev, launched in September. Both models belong to a growing category called system-one models—fast, routine decision-makers designed to work alongside larger language models rather than replace them. OpenAI's recent Decisions API preview signals the same strategic direction across the industry.

Strands Decider 2B is integrated into AWS's Strands Agents program, which bundles open-source decision-making capabilities into the AWS compute ecosystem. The move reflects a clear capital allocation bet: lower the total cost of ownership for AI agent deployment, expand the addressable market for AWS inference workloads, and compete on developer momentum.

The pointer-head architecture has inherent limits. It functions only on closed-choice problems—tasks requiring open-ended or generative answers still need a traditional language model. Its second-place ranking on JevBench's full benchmark, and perfect score only on the easy tier, leaves questions about real-world performance on complex decision scenarios where true competition with Jev will surface.