Amazon Releases Strands Decider 2B, an Open-Source Decision Model to Accelerate Agentic Workflows

AWS Enters the Decision Model Race with Open-Source Offering

Amazon Web Services released Strands Decider 2B, an open-source decision model designed to accelerate agentic workflows by delivering high-speed, low-cost choices between predefined options1

1

. The release positions AWS alongside TypeSafe and other developers in the emerging decision model space, where AI systems choose among structured options rather than generating text. Marc Brooker, an Amazon distinguished engineer, conceived the project after encountering TypeSafe's Jev model and building his own implementation that briefly topped the JevBench ranking for models of its size1

1

.

The model addresses a specific need AWS customers identified: agentic workflows don't always require the capability or cost of fully-featured LLMs. Strands Decider 2B delivers calibrated choices with confidence scores, making it suitable for workflow steps that need to determine the next action based on current state1

1

. Available now on Hugging Face under an Apache 2.0 license, the model is free to download with complete training materials, scripts, and examples on GitHub2

2

3

3

.

Technical Architecture Balances Speed and Intelligence

Strands Decider 2B builds on the Qwen3.5-2B base model but replaces the traditional text-generation component with a customized pointer head containing just over 1 million parameters2

2

. AWS adapted the base model with a rank-16 LoRA update and trained the new scoring component to evaluate answer options directly in a single forward pass2

2

. This architecture, labeled "Hobson" in AWS documentation, scores supplied answer choices without generating explanatory text, dramatically reducing latency and computational costs3

3

.

The 2 billion-parameter scale represents a deliberate choice by AWS engineers. This size provides sufficient power for complex decisions while remaining small enough to run locally with less than 150 milliseconds of latency3

3

. Testing showed median latency of 106 milliseconds and 95th-percentile latency of 296 milliseconds across 230 requests on an Nvidia RTX 3090, including HTTP round trip2

2

. AWS also reported roughly 150 milliseconds median for small tasks on an M3 MacBook2

2

.

Performance Metrics and Competitive Positioning

On JevBench testing, Strands Decider 2B version 19 achieved roughly 72% accuracy with a 0.35 Brier score, ranking second among public models around its size and first among those publishing complete training recipes2

2

. The similarly sized Mapika decider-2b v11 scored around 76% accuracy with a 0.32 Brier score, suggesting room for improvement in AWS's offering2

2

. AWS has iterated rapidly on the model, with today's release representing version 202

2

.

Brooker emphasized the challenge lies in optimizing rapid decision-making without compromising intelligence. "There is a very careful balance to be found where you want to push its performance on accuracy and calibration on these kinds of tasks, without degrading its performance on understanding different languages, on having the kind of knowledge it has, which is what makes it general purpose and interesting and useful," he explained1

1

. He noted that smaller market sizes mean building competitive models costs only hundreds or thousands of dollars, potentially democratizing development beyond frontier labs1

1

.

Practical Applications in Agent Workflows

The model integrates into agentic workflows for tool selection, model routing, context management, guardrail enforcement, and policy classification3

3

. AWS's demonstration places Strands Decider 2B before an agent calls a weather tool, examining whether a city came from the user's request and whether it's too early to proceed with the tool call2

2

. The application code uses these answers to either proceed or send the agent back for clarification2

2

.

This approach enables hybrid agents that use decision models for simple, repetitive choices while reserving LLMs for complex reasoning challenges3

3

. The confidence scores assigned to each decision help developers set appropriate thresholds for proceeding, denying tool calls, or requesting confirmation2

2

. AWS acknowledged that dedicated decision-model integration libraries remain in development, with current implementations requiring developers to define how applications handle the model's judgments2

2

.

Market Response and Future Outlook

TypeSafe CEO Diogo Almeida responded to the wave of competing models by suggesting developers may be underestimating the difficulty of making models truly intelligent. "The current batch seems more like ML people wanting to implement a cool architecture than a team deeply dedicated to making intelligence useful," he stated1

1

. He indicated TypeSafe would focus on improving future models rather than reacting to competition1

1

.

The proliferation of decision models since TypeSafe's debut demonstrates wide interest in cost efficiency and low latency for AI systems. Watch for continued optimization battles as developers balance speed against accuracy and calibration. The open-source nature of Strands Decider 2B, combined with complete training recipes, may accelerate experimentation and reveal whether decision models become standard components in production agent systems or remain specialized tools for specific workflow checkpoints.