Amazon Web Services has released an open-source decision model called Strands Decider 2B, a small system that sorts between pre-decided options and reports how confident it is in each choice rather than generating prose. TechCrunch reported the release in the same week that OpenAI announced a similar offering, and described the wider field of decision models as flooding the web.
According to the report, this model is completely open-sourced, can be run locally because of its small size, and is available right now. Rather than generating text, it provides calibrated choices, and it is constructed on the "torso" of an already existing language model — Qen3.5-2B in this instance. That architecture is not exclusive to Amazon's release; it is shared with other decision models.
Marc Brooker, a distinguished engineer at Amazon, started the project as a homebrew effort after seeing TypeSafe's Jev and attempting to create his own version. Per the report, the side project succeeded enough to briefly reach the number one position on the Jevbench ranking for models of its size. Following that, Amazon engineers cleaned it up and released it through Strands Labs, an organization that develops tools and protocols for deploying AI agents.
Brooker told TechCrunch that AWS customers' agentic workflows had driven the need, since those workflows did not always call for the capability or cost of a fully featured LLM. He described this class of model as well suited to a workflow step that answers which action should come next given the current state of the task — a decision that, he said, can be structured more reliably because of the confidence scores and the closed domain of answers, with lower latency and potentially lower cost.
The naming lineage runs back to TypeSafe, which called its model Jev after the economist William Stanley Jevons, invoking his theory that falling costs for something like computer intelligence can increase demand for it. Since TypeSafe debuted the idea, the report says dozens of similar models have been produced by researchers, which demonstrates wide interest but also raises the question of how valuable any individual model in the category can be.
Brooker framed the central engineering tension as a balance between optimizing fast decision-making and preserving general capability. He told TechCrunch that performance on accuracy and calibration for these tasks has to be pushed without degrading performance on understanding different languages or on retaining the kind of knowledge that makes a model general purpose, interesting and useful.
He also said he does not necessarily expect frontier labs to dominate the space, particularly because the markets are smaller and the cost to build something interesting sits in the hundreds or thousands of dollars. That cost estimate is his characterization of the category, not a published benchmark or a figure Amazon has attached to Strands Decider itself.
TypeSafe executives, for their part, said they are keeping their heads down and improving future models. Chief executive and founder Diogo Almeida told TechCrunch that people may be underestimating how hard it is to make these models actually smart, and that he did not yet see real competition emerging for his company. He characterized the current batch of entrants as machine-learning practitioners wanting to implement a cool architecture rather than teams deeply dedicated to making intelligence useful.
For freelancers and developers building agent workflows, the practical significance is the shape of the tool rather than its size. A model that returns a bounded set of choices with a confidence measure can be dropped into a single step of a pipeline — routing, triage, or picking the next action — where a full LLM would be slower and more expensive. Because it is open-sourced and small enough to run locally, it can be evaluated without committing to a hosted API, though the report gives no pricing, licensing terms, or hardware requirements.
The confidence score is the part that matters most for production use, and also the part that is hardest to trust without testing. A calibrated score is only useful if it is actually calibrated on your task and your data; the report describes the model as delivering calibrated choices but does not present accuracy figures, evaluation methodology, or independent verification of that calibration. Anyone adopting it should treat the confidence output as a claim to be measured against their own decision set.
The competitive context cuts both ways. On one side, dozens of researchers have already produced similar models, which suggests the approach is easy enough to replicate and that differentiation may be thin. On the other, Brooker's argument that small markets keep build costs low implies the category may stay crowded with cheap alternatives rather than consolidating around a few vendors. Almeida's counter-argument — that implementing the architecture is not the same as making the model smart — is the strongest stated reason to be skeptical of the flood.
There is also an unresolved question about what happens to general capability as these models are tuned for speed and calibration. Brooker's own description of the tradeoff implies that pushing accuracy on narrow decision tasks can come at the cost of language understanding and stored knowledge. That is a design constraint worth watching for anyone who wants one small model to handle both structured decisions and the messier language work around them.
What the evidence does not establish is how Strands Decider 2B performs relative to Jev, to OpenAI's comparable offering, or to the other models in the field. The report cites a brief top placement on Jevbench for models of its size during the homebrew phase, but gives no current ranking, no head-to-head results, and no detail on what Jevbench measures. The release date is also not pinned to a specific day in the supplied material; it is described only as the same week as OpenAI's announcement.
For a working developer, the sensible reading is that this is a low-commitment option to test rather than a settled replacement for existing tooling. The absence of published pricing, benchmarks, and licensing detail means the decision to adopt should rest on local evaluation against real workflow steps. The category's direction — cheap, small, locally runnable decision models — looks durable; which specific model wins that role does not yet.