AWS Open-Sources Strands Decider 2B, a Decision Model That Picks in Under 100 Milliseconds
October 2, 2026 — Amazon Web Services wants your AI agents to stop writing when all they need to do is choose.
What happened
On October 1, 2026, AWS’s Strands Labs released Strands Decider 2B, an open-source “decision model” designed to sit inside an agent’s workflow and decide whether a proposed action should proceed. The weights are free to download from Hugging Face under the permissive Apache 2.0 license, with training data and scripts published on GitHub.
How it works
The idea is the same one driving this week’s rush of decision models: when a program only needs an answer like “Should this tool run?” or “Which of these three routes fits this request?”, asking a large language model to write an explanation adds time and expense. Strands Decider 2B scores the allowed answers directly, in a single pass, and returns each choice with a confidence score. AWS built it on a Qwen3.5-2B base model and fine-tuned it with LoRA, stripping out text generation entirely in favor of what AWS calls a “pointer head” — a roughly 1-million-parameter add-on that looks at the predefined options and scores them without generating words.
The headline is speed: AWS says the model makes decisions in under 100 milliseconds on widely available hardware, including an NVIDIA RTX 3090, making it suitable for running locally on a CPU or GPU. On JevBench, the community benchmark for decision models, it posted a perfect score on the easy tier and ranks among the top public models of its size class — first among those shipping the full training recipe.
A side project that grew up
The backstory is almost as interesting as the model. Marc Brooker, an AWS distinguished engineer, started the project at home after seeing Typesafe AI’s Jev launch in September; his homebrew version briefly topped the JevBench ranking for its size class. AWS engineers then cleaned it up and shipped it through Strands Labs, the company’s group for agent tooling. Brooker said the push came from AWS customers whose agent workflows didn’t always need the capability — or the cost — of a full LLM for every step.
Why it matters
Decision models are becoming the guardrail layer of the agentic stack: cheap, fast checkpoints that route requests, select tools, verify an agent’s arguments, or gate premature tool calls before the expensive flagship model does anything irreversible. Confidence scores are the crucial feature here — without them, a model that can’t explain itself is just a black box saying “trust me.”
And openness is the real differentiator. Weights, training data, and scripts all published means any team can run this on their own hardware, forever, with zero per-call cost — a meaningful contrast to closed-API alternatives. For developers, the practical move is the hybrid agent: a local decider for the routine gates, a full LLM for the hard reasoning. AWS just handed you the decider for free.
Frequently asked questions
What can I use Strands Decider 2B for? Routing requests, selecting tools, evaluating outputs, and reviewing an agent’s actions — anywhere an agent needs a quick, bounded yes/no or multiple-choice decision.
Can it run on my laptop? Yes, that’s the point. At 2 billion parameters it’s small enough for local deployment on widely available hardware, including an RTX 3090, returning decisions in tens of milliseconds.
Is it actually open source? Yes. Apache 2.0 license, weights on Hugging Face, training data and scripts on GitHub.
How does it compare to Typesafe’s Jev? AWS reports it matched or outperformed Jev and other decision models on accuracy and confidence calibration in its evaluations, placing second among public models of roughly its size.
Sources: VentureBeat, TechCrunch, Crypto Briefing, SiliconANGLE