Strands opens Decider 2B for bounded agent decisions

Strands released Decider 2B on Thursday, making its code, weights, training data and training scripts available for download. The roughly two-billion-parameter model scores supplied options or assigns a score; it does not generate prose.

For developers, the release creates a self-hosted layer for frequent decisions such as routing a request, choosing a tool or checking proposed tool arguments without sending every step to a larger generative model. It sits alongside Strands’ open-source agent harness for local and cloud use. Strands lists evaluations, guardrails, policy classification and hybrid agents among the model’s intended uses, while leaving writing, coding and complex reasoning to generative models.

Decider 2B starts with the Qwen3.5-2B-Base decoder torso. Strands removed the language-modeling head that predicts text and replaced it with a pointer head of about one million parameters. In one forward pass, that head compares internal representations and scores the answer choices supplied with a question. A rank-16 LoRA adapter modifies the base model for the task. The public repository lists the more precise size as 1.9 billion parameters, making “2B” a rounded model name.

The repository contains source code and directories for data, evaluation, research and training, along with recipes for local and distributed training. The project is licensed under Apache 2.0. Its v19 checkpoint is available from Hugging Face, so developers can inspect or adapt the release without relying on a Strands-hosted inference API.

The documented local path starts with installing the strands-decider Python package and calling the Hugging Face checkpoint from the command-line interface. The software runs with CUDA, Apple MPS or a CPU and can expose an unauthenticated loopback HTTP endpoint for local experiments. Users must supply and operate the compute.

In Strands’ worked agent example, the decider runs immediately before a weather-tool call. If an agent proposes a city the user did not provide, the model scores two yes-or-no questions: whether the arguments are grounded and whether the call is premature. Application code then returns control to the agent so it can ask for the missing city. Strands says the example’s questions, threshold and policy were selected by hand, so a model score alone does not determine the application’s response.

Strands says the v19 model answered 167 of 231 public JevBench tasks correctly, for an accuracy of 0.723. Strands also reports a median of 115 milliseconds per JevBench question on an RTX 3090 under its test setup. Neither result has been independently reproduced in the launch evidence.