AWS open-sources Strands Decider 2B for fast local agent decisions

Amazon Web Services released Strands Decider 2B on Oct. 1, an open-source “decision model” meant to choose among predefined options in tens of milliseconds — a cheaper, more structured step for agent workflows than calling a full large language model every time.

Announced on the Strands Agents blog by AWS distinguished engineer Marc Brooker and colleagues, the roughly two-billion-parameter system starts from Alibaba’s Qwen3.5-2B torso, strips the text-generation head and adds a small scoring head so it returns calibrated choices and confidence scores in one forward pass. Weights are on Hugging Face and training code is on GitHub under Apache 2.0; AWS is not offering a hosted pay-per-call API. Brooker told TechCrunch the idea grew from customer agent pipelines that needed reliable next-step routing — tool selection, guardrails, model routing — without frontier-model latency or cost. On a local Nvidia RTX 3090, AWS reports median decision latency around 115 milliseconds for small tasks; an M3 MacBook landed near 153 milliseconds in their tests.

Decision models surged after TypeSafe’s Jev drew attention earlier this year. TechCrunch noted Amazon’s release landed the same week OpenAI floated a similar idea, while TypeSafe CEO Diogo Almeida argued many clones still lag on real intelligence. Strands Labs positions Decider 2B as an experiment-friendly baseline: small enough to fine-tune on ordinary hardware, competitive on the public JevBench leaderboard for its size class, and deliberately unsuitable for chat, coding or long-form generation.

For Canadian developers and enterprises already wiring agent stacks, a free local decider lowers the cost of putting a confidence-gated checkpoint before tool calls — provided teams accept the ops burden of self-hosting and treat scores as workflow signals, not guarantees.

Sources: Strands Agents / AWS Strands Labs; TechCrunch.