Strands Decider 2B: AWS Open-Sources a 115ms Agent Router

AWS Strands Labs has released Strands Decider 2B, an open-weight model built for one narrow job. It makes the small, repeated choices an AI agent faces during a task, and it does this without generating any text. The model arrived on October 1, 2026, under an Apache 2.0 license 12. It extends AWS's open-source Strands Agents framework, which has been maturing steadily since its 2025 debut 1.

The release matters less for its size than for its design. Agent builders have long used general-purpose language models for decisions that don't need prose. Decider tries to replace that habit with a dedicated, cheap component that can be run locally.

What AWS shipped

Decider is a decoder model of roughly 1.9 billion parameters. Its standard language-modeling head has been swapped for a lightweight "pointer head" that adds about a million parameters 1. The model does not predict the next token. It reads the current situation, compares its internal hidden-state representation against each candidate option, and returns a ranked list of probabilities 2.

According to benchmarks included in the package release, that ranking comes back in a median of 115 milliseconds on a consumer GPU 2. Both outlets use that figure as the model's headline number 12.

The weights are on Hugging Face at StrandsAgents/strands-decider-2B-hobson-v19 2. The project also ships through PyPI and installs with a single pip install strands-decider command. It includes a command-line interface and an HTTP server, which makes it easy to run as a standalone routing service 2.

The framework around it

Decider is designed to sit inside the Strands Agents SDK. The SDK is AWS's open-source, model-driven framework for building agents in Python and TypeScript, and it is now at version 1.59.0 1. Its history:

• It launched as a preview in May 2025 1.
• It reached a production-ready 1.0 release about two months later 1.
• AWS reported more than 2,000 GitHub stars and 150,000 PyPI downloads in that early period 1.

That version number points to a fast release cadence. AWS has been treating Strands as an actively developed platform rather than a one-off demo. Decider looks like the next layer on that base, a purpose-built part for the agent loop rather than a general model squeezed into it.

Why a non-generative decision model matters

An agent working through a task makes many small decisions: which tool to call, which branch to take, whether a step is finished. The tech-insider coverage notes that these choices can happen dozens of times per task 1. Sending each one to a large hosted model adds latency, cost, and a dependency on an external API. Free-form text output also has to be parsed back into a structured choice, which can fail.

Decider addresses this by never producing text. Because it scores a fixed set of candidates, its output is always one of the valid options, with a probability attached 2. The model cannot ramble or return a malformed answer, and the ranked probabilities give developers a confidence signal they could use to fall back to a larger model when the decision is close. That fallback pattern is our inference about how teams might use it, not a feature AWS has documented.

How open is it?

The two reports frame the release a little differently. Tech-insider presents Decider mainly as a companion to the Strands SDK and Bedrock-based agent development 1. TechTimes emphasizes openness. It calls Decider the first fully open-source entry in an emerging "decision-model" infrastructure category, because the weights, training recipe, and evaluation harness are all public 2. It also says this lets developers outside the reach of hosted APIs self-host the routing layer 2.

The "first" claim is hard to verify independently, and the category itself is still loosely defined. But publishing the training recipe and evaluation harness is a stronger commitment than releasing weights alone. It lets others reproduce the results, audit the 115ms latency figure, and fine-tune the model for their own domains. The Apache 2.0 license allows commercial use, modification, and redistribution without licensing fees 2.

Our take

The sources agree on the core facts: the release date, the license, the 115ms median latency, and the pointer-head architecture 12. Where they differ is emphasis. Tech-insider places Decider within AWS's agent tooling, while TechTimes focuses on its potential to be used independently 12.

Both framings are probably true, and the tension between them is the interesting part. AWS benefits when developers build on Strands and Bedrock. Yet it has released a component that is useful well outside that ecosystem. If small decision models prove reliable, agent design could shift toward a split approach: a large model for reasoning and generation, and a tiny, fast scorer for the frequent routine choices. The real test is whether Decider's accuracy holds up on messy, real-world agent tasks, not just its speed.