Strands Decider 2B launches for local AI agent decisions

Strands introduced Strands Decider 2B on October 1, releasing a small decision model for local AI agent experiments. The launch announcement describes a model that chooses among supplied options or assigns numerical scores. Its intended uses include model routing, tool selection and policy classification.

The design gives up text generation to focus on decisions. It cannot write a chat response, summarize a document or generate code. The Strands team positions it as a component in workflows where larger language models handle more demanding reasoning and a smaller model handles routine choices.

How the smaller model works

According to the project repository, the model has 1.9 billion parameters. It starts with the Qwen 3.5 2B base model, removes the language modeling head and adds a pointer head of roughly one million parameters. A LoRA adapter adjusts the underlying model. The pointer head scores the choices supplied in each request through a single forward pass.

The published model card identifies v19 as an adapter and readout head distributed under Apache License 2.0. The base weights download separately on first use. The card links to the training recipe, data inventory and evaluations, giving developers materials for reproducing or adapting the model.

Published results come with limits

For the released checkpoint, the model card reports 167 correct answers across 231 public JevBench tasks, about 72.3 percent accuracy. It lists that same count at context windows of 3,072 and 4,096 tokens. These are results reported by the model publisher, rather than an independent test by ByteForward.

The project’s results record separates the original reference run from a later retrain on AWS hardware. The two runs differ on individual benchmark tasks despite matching the overall count at 3,072 tokens. That is a useful warning against treating every v19 result as a measurement of identical weights.

The evaluation notes also describe substantial weaknesses. In one question sensitivity test, changing the question while holding the state and options fixed left v19’s answer unchanged about 94 percent of the time. Negated instructions can be missed, while long reasoning tasks and unfamiliar scoring rubrics remain difficult.

Confidence scores need similar care. The documented confidence bands were established on short classification tasks, and the authors advise testing them on an application’s own traffic before trusting an automation threshold. The notes also caution that the small public benchmark cannot reliably settle modest differences between individual training runs.

Local testing before broader deployment

The inference guide documents command line use and an HTTP server. Serving is supported on Nvidia GPUs, Apple silicon and CPUs, with CPU inference substantially slower. Multiple questions can share the same input text through caching, reducing repeated processing.

The server is intended for local experiments. It binds to the local machine by default, has no authentication and has not been verified under concurrent requests. Those limits matter for anyone considering moving a demonstration into a shared application.

The repository still names v19 as the reference model. A later v20 experiment did not meet its own promotion criteria and did not replace it. Developers evaluating the release should therefore distinguish the published v19 checkpoint from experimental results elsewhere in the project.

For a team trying it, a useful first test would be a narrow routing task with clear labels and human review of uncertain decisions. The important question is whether it handles that workload reliably enough to justify a separate decision component.

Our coverage of Cloudflare Clef and Clef Flash examines a larger decision model family with hosted inference and its own quality tradeoffs.

Original AI generated conceptual illustration of a compact local model selecting a bounded choice