Strands Releases a Small Open Source Model That Cannot Write Text
Strands Agents released Strands Decider 2B on Oct. 1, 2026, publishing the model's weights, its code, its training data, and the scripts used to train it the same day. The team describes it as a decision model: it picks between the options it is offered and scores them, and it cannot generate text at all.
The release post says decision models are faster at a given size than text-generating models, always return one of the options they are given, and attach a reliability score to every decision. The team says that combination suits the kind of work agents do: routing a query to the right model, choosing a tool, applying guardrails, and sorting text against a policy. The post also says the same design rules the model out for coding, chatbots, and document summarization.
The architecture starts from an existing open language model, Qwen3.5-2B, and strips out the part that produces words. In its place sits a scoring head of just over a million parameters, which rates the options the model is given and returns a confidence number with its pick. The whole model carries 2 billion parameters, and the team says it is small enough to run on a local CPU or graphics card.
On its own measurements, the team places the model third of 33 in its size class for accuracy, and for how trustworthy its confidence scores are, both scored on the public half of a benchmark called JevBench. The team also reports a median decision time of about 115 milliseconds on a desktop Nvidia RTX 3090 and about 153 milliseconds for small tasks on an M3 MacBook. All of those are the team's own figures on its own model, and no independent evaluation of them has been published.
