AWS open-sources Strands Decider 2B, a millisecond decision model for agent routing

Strands Decider 2B belongs to what AWS calls a new class of decision models. A reasoning model generates open-ended text. A decision model instead chooses among options you provide and attaches a confidence score to each. Swami Sivasubramanian, AWS VP of Agentic AI, described them as faster and more capable at a given size, and noted they always return an answer from the allowed set. The 2-billion-parameter model runs on a local CPU or GPU and answers in tens of milliseconds. Reports put latency at about 115ms on an RTX 3090, and Hugging Face listings cite 106ms. It scores 0.723 on JevBench.

Matt Wood, AWS's Chief AI & Technology Officer, listed the intended uses. These include simple classifiers such as spam detection or sentiment, plus the routine decisions an agent makes on every turn: which tool to call, which model to route a request to, whether an action passes a guardrail, and whether a request fits a policy. The model is fine-tuned from Qwen 3.5-2B. AWS released both the weights and the full training recipe under Apache 2.0, so teams can retrain it on their own decision data.

The competitive target is clear. VentureBeat framed it as a free, fast "Jev killer," aimed at TypeSafe's Jev decision model. Perplexity moved into the same space the same day, open-sourcing pplx-decider-27b with a Decisions API priced at 4 cents per million input tokens and free output. Decision models are suddenly a crowded category, and open weights are commoditizing them before any one vendor can lock in the market.

The caveats are about scope. A 2B model cannot do open reasoning, and calibration quality on out-of-distribution decisions is what will decide whether teams trust it as a guardrail. JevBench is also a young benchmark. The thing to watch is whether Strands Decider becomes the default router inside AgentCore and Kiro workflows, which would give AWS a sticky, cheap layer inside customers' agent stacks.