Strands Agents Releases Strands Decider 2B, an Open Source Decision Model

Strands Agents has released Strands Decider 2B, a small open source decision model designed for fast experimentation and local development. The 2 billion parameter model can answer meaningful questions in tens of milliseconds on widely available hardware, and the team has published the code on GitHub along with model weights, training data, and scripts on Hugging Face.

Decision models versus LLMs

Strands Decider belongs to a new class of models sometimes called decision models or system one models. Unlike LLMs, which generate arbitrary text, these models pick between sets of options, such as classifying a phrase as English, Zulu, or Dutch, or scoring sentiment between 0 and 1. In exchange for reduced flexibility, they are faster and more capable at a given size, always return an answer from the selected options, and run with very low latency.

The tradeoff is that generating all outputs in a single parallel pass makes them weaker than reasoning models on complex problems, and their inability to generate text rules out coding, chatbots, and document summarization. On the other hand, each decision comes with a reliability score that frontier LLM inference APIs do not provide, and the models can answer multiple questions about the same prompt highly efficiently.

Architecture and performance

The model starts from a pre-trained Qwen3.5-2B torso with the language modeling head removed, taking away its ability to generate text. A pointer head of just over one million parameters scores the hidden state at each option position against the hidden state at the answer position, and the torso is fine-tuned with a rank-16 LoRA adapter. The released model is version 19; an earlier architecture using a slot head performed significantly worse, and the repository documents every change across iterations.

On JevBench's public set, Strands Decider 2B ranks third of 33 models in the 2B class for accuracy and calibration measured with the Brier score, and first of 30 when just-over-2B models are excluded. Median decision latency is around 115 milliseconds on a local Nvidia RTX 3090 and about 153 milliseconds for small tasks on an M3 MacBook, increasing approximately linearly with task size. The model answers 100 percent of the easy tasks on JevBench correctly.

Getting started

The easiest entry point is the strands-decider CLI, installed with pip install strands-decider. A strands-decider ask command takes a state, a question, and a set of choices, and returns probability scores plus a confidence value for each option. The repository also includes an example of the model running inside a Strands agent, where it checks whether a proposed tool call is grounded in user input and whether it is premature, using the framework's intervention system to block or guide the call before it executes.

Early use cases reported by the team include model routing, tool selection, evaluations, guardrails, memory, context management, and policy classification, as well as hybrid agents that use LLMs for hard decisions and decider models for routine ones. The Strands team says it is working on libraries for decision model integration.

Sources: strandsagents.com