Summary
Strands Labs has released Decider 2B, a 2-billion-parameter model designed to choose among supplied options and provide confidence scores. The team reports JevBench results and local latency measurements, and has published the model, training data and scripts.
Strands Labs announced Strands Decider 2B on October 1, 2026: an open-source, 2-billion-parameter model built to make decisions from a set of options rather than generate unrestricted text. The team says it has published the model on GitHub and its weights on Hugging Face, along with the training data and scripts.
The model is intended for tasks such as selecting a tool, routing a request or classifying a policy question. Strands Labs reports that it ranked third among 33 models in the 2B class on JevBench’s public set, and first among 30 when models just over 2B parameters were excluded. The team also reports a median local decision time of around 115 milliseconds on widely available hardware.
How a decision model works
A language model typically generates text token by token. Decider 2B instead scores options supplied with a question or task state. For example, an agent could ask whether a proposed tool call is appropriate, or choose which team should handle a request from a list of departments. Its output is constrained to the available choices and includes scores that can be used as confidence estimates.
That narrower task can make a small model useful in parts of an AI system where a full text-generating model would be unnecessary. Strands Labs says decision models can make fast, repeated choices and provide reliability scores. They are not designed for tasks that depend on generating extended text, such as coding, chat or document summarisation, or for complex reasoning.
Model design and reported results
Decider 2B uses the pre-trained Qwen3.5-2B model as its base. Strands Labs says it removed the language-model head—the component used to generate text—and replaced it with a pointer head. The replacement scores the hidden representations of the offered options against the representation at an <answer> position. The pointer head has just over one million parameters, and the model’s main component is fine-tuned with a rank-16 LoRA adapter, a method for adapting a model by training a smaller set of added parameters.
The team assessed accuracy and calibration on JevBench’s public set. Calibration concerns how well confidence estimates correspond to actual outcomes; the Brier score is one measure of that relationship, with lower scores indicating better-calibrated probability estimates. Strands Labs reports Decider 2B’s accuracy and calibration ranking as third of 33 in the 2B class, or first of 30 excluding models just over 2B. It also says the model answered 100% of JevBench’s easy tasks correctly.
For latency, the post reports a median of around 115 milliseconds on widely available hardware and around 153 milliseconds for small tasks on an M3 MacBook. It says task time rises approximately linearly with task size. The post’s latency graph specifies a local Nvidia RTX 3090 and version 18 of the model, so those graph measurements refer to that version rather than the released v19 model.
Using it inside an agent
One example pairs the model with a Strands agent that can call a weather tool. Before the tool runs, Decider 2B evaluates whether the proposed location came from the user and whether the call should wait for clarification. The agent’s Python intervention handler uses those answers to proceed, block the call, request confirmation or return the conversation for clarification.
The example shows a possible role for a fast decision model: checking routine choices around a larger language model, rather than replacing it. Strands Labs describes the questions, threshold and policy in this demonstration as hand-picked, making it an illustration of the integration pattern rather than a general-purpose safety guarantee.