Amazon releases Strands Decider 2B open decision model

Amazon Web Services' Strands Labs released Strands Decider 2B, an open-source decision model, on 1 October 2026. The model sorts between pre-decided options and returns a measure of how confident it is in its choice, and it is small enough to run locally. It arrived the same week OpenAI announced a similar offering, as AI developers seek intelligence better suited to computer automation than frontier large language models.thenewstack.io+1

Strands Decider matters because agentic workflows — automated sequences where an AI agent picks its next action — do not always need the capability or cost of a fully featured large language model at every step. Strands Decider offers a workflow step that is more reliable through confidence scores and a closed domain of answers, with lower latency and potentially lower cost, Brooker said.

What the model actually does

The model is built on the torso of an existing large language model, in this case Qwen3.5-2B, but instead of generating text it delivers calibrated choices. A decision model makes classifications to help agents decide how to act based on certain probabilities, returning typed answers with probabilities that code can use to route work or defer to a human.

Amazon distinguished engineer Marc Brooker originated the project after seeing TypeSafe's Jev and trying to build his own take on such a model. The homebrew version was successful enough that it briefly reached the top spot on the Jevbench ranking for models of its size, after which Amazon engineers cleaned it up and released it through Strands Labs.

Brooker said the need emerged in conversations with Amazon Web Services customers whose agentic workflows did not always require the capability or cost of a fully featured large language model all the time. The model is fully open sourced and available now, released the same week OpenAI announced a similar offering, as dozens of similar models have appeared since TypeSafe debuted its idea.

Where the numbers came from

TypeSafe named their model Jev after the economist William Stanley Jevons, hoping to invoke his theory that the falling cost of something, like computer intelligence, can in fact increase its demand. Cloudflare released its own decision models, Clef and Clef-flash, hosted on Workers AI under an Apache 2.0 licence, and Clef is currently the leader when evaluated against the Jev Decision Index.techcrunch.com

Brooker described the central engineering trade-off as a careful balance between pushing performance on accuracy and calibration without degrading performance on understanding different languages and retaining the knowledge that makes a model general purpose. He does not necessarily expect the frontier labs to dominate the space, noting that with smaller markets the cost to build something interesting is in the hundreds or thousands of dollars.

The trade-off Brooker describes

Cloudflare's own testing illustrates the latency argument for decision models. Its Threat Intelligence team used Clef to classify a website domain, and the model took 2.2 seconds to fetch, render and classify the site, returning probabilities such as 95% fashion and 85% ecommerce. Cloudflare's fastest general large language model, gpt-oss-120b, took 4.7 seconds in the same workflow and returned only two classifications.

TypeSafe chief executive and founder Diogo Almeida said he did not see real competition for his company emerging yet, and that his team is keeping its heads down improving future models. He said the current batch seems more like machine learning people wanting to implement a cool architecture than a team deeply dedicated to making intelligence useful.

Competition in a crowded field

The open question is how valuable the dozens of similar models can be. Brooker suggests the challenge will be in optimising the model's speedy decision-making without compromising its intelligence, while Almeida said people might be underestimating the difficulty of making the models actually smart. Cloudflare, meanwhile, debuted a reinforcement learning platform that lets customers fine-tune Clef on their own data.