Strands Decider 2B

Not yet independently verified. Every detail here comes from the Strands Agents project’s own announcement and its Hugging Face repository; the only independent report found (VentureBeat, 1 October 2026) restates it. Accuracy and calibration claims are the project’s own. We will update this when it can be confirmed, and remove this note.

Strands Decider 2B is a small open-source decision model released on 1 October 2026 by the Strands Agents project, AWS’s open-source agent framework, in a post by Marc Brooker, Mike Chambers and Fabio Nonato de Paula. Instead of generating text, a decision model chooses among options supplied with the prompt (yes or no, one of several languages, a category) or returns a numeric score, and attaches a reliability figure to each answer; the authors say this makes it fast and dependable for routing and checks inside agent workflows, and unsuited to coding, chat or summarisation. Technically it takes the Qwen 3.5 2B base model, removes its text-generation head and adds a small pointer head that scores each option, with the base fine-tuned through a LoRA adapter; the released version is the nineteenth iteration. The project says it runs on a local CPU or GPU and answers in tens of milliseconds. The weights are on Hugging Face under the Apache 2.0 licence, and the code, training data and scripts are on GitHub, so the result can be reproduced. The authors describe its accuracy and calibration as competitive with other decision models; those are their own measurements, so no score is stored. The board also lists the Clef decision models released the same day.