AWS Launches Strands Decider 2B: An Open-Weight Local Decision Model for AI Agents

Amazon Web Services has entered the fast-growing category of specialized agent decision models with the release of Strands Decider 2B, an open-weight 2-billion-parameter model accompanied by its complete training dataset and fine-tuning scripts. Rather than generating free-form conversational text, Strands Decider 2B evaluates an agent’s current state against a developer-supplied list of options or safety checks, returning calibrated probability scores in under 100 milliseconds on a single desktop GPU.

Understanding the topic starts with recognizing a common inefficiency in modern AI engineering. When an autonomous agent needs to make a binary or multiple-choice routing decision—such as verifying whether a user actually supplied a required city parameter before calling a weather API, or deciding whether to escalate a support ticket to a human—most teams invoke a full generative large language model. Paying a frontier model to generate tokens just to output “yes” or “no” introduces unnecessary network latency, unpredictable token bills, and the risk that the model hallucinates an invalid option outside the schema.

Decision models solve this architectural bottleneck by stripping away open-ended text generation altogether. Incubated inside AWS Strands Labs alongside the company’s open-source Strands agent framework and Strands Harness runtime, Strands Decider 2B uses Qwen3.5-2B as its language-understanding torso paired with a rank-16 Low-Rank Adaptation (LoRA) adapter. AWS engineers removed the standard causal language-modeling head and replaced it with a compact pointer head containing roughly 1 million parameters that directly scores the developer’s candidate choices.

Because the pointer head can only assign probabilities to the exact options passed in the request, the model is structurally incapable of inventing an unapproved action. On the public JevBench evaluation suite—originally created by TypeSafe AI to benchmark its proprietary Jev decision API and OpenAI’s hosted Luna Decisions API—Strands Decider 2B ranks second among all 2-billion-parameter public models, ranks first among models that publish a reproducible training recipe, and achieves 100% accuracy on JevBench’s easy tier for routine agent routing.

As reported in The New Stack’s analysis of AWS Strands Decider 2B, the model runs locally alongside cloud agents calling Amazon Bedrock, completing small routing checks in under 100 milliseconds on an NVIDIA RTX 3090 and roughly 150 milliseconds on an Apple M3 MacBook Pro. By publishing every earlier architectural iteration in the repository, AWS gives engineering teams a transparent blueprint for training custom in-house gatekeepers without relying on third-party decision APIs.

Fast Facts: AWS Strands Decider 2B Architecture & Performance

• Base Model Torso: Built on open-weight Qwen3.5-2B using a rank-16 LoRA adapter and a custom 1M-parameter pointer head instead of a text-generation head.
• Core Function: Evaluates conversation state and proposed tool arguments against predefined developer options, returning calibrated confidence scores.
• Inference Latency: Under 100 milliseconds on a single NVIDIA RTX 3090 GPU and approximately 150 milliseconds median latency on an Apple M3 MacBook.
• Benchmark Standing: Ranked #2 among ~2B-parameter public models on JevBench (#1 among models with open training recipes) and 100% on JevBench easy tier.
• Open-Source Release: Includes model weights, training scripts, calibration datasets, and historical architecture iterations via AWS Strands Labs.

How Strands Decider 2B Compares to Competing Decision Models

How Local Pre-Tool Verification Prevents Agent Hallucinations

In AWS’s reference architecture, Strands Decider 2B sits directly inside the Strands intervention hook between the generative reasoning model (hosted on Amazon Bedrock) and external enterprise APIs. When a user asks a vague question and the primary LLM attempts to guess missing tool parameters, Strands Decider intercepts the proposed tool call locally in 100 milliseconds, verifies whether every argument is explicitly grounded in the transcript, and bounces the turn back to ask the user for clarification before an erroneous API call ever fires.

The 4-Step Enterprise Implementation Playbook

1. Insert a Local Decision Gate Before Mutating Tool Calls: Deploy Strands Decider 2B as a local sidecar container on your agent worker nodes to verify that every proposed database write, API POST, or customer email has grounded parameters before execution.
2. Use Calibrated Confidence Thresholds for Human-in-the-Loop Escalation: Because AWS trained Strands Decider 2B specifically for probability calibration, configure your agent router to auto-execute actions scoring above 0.90, re-prompt the agent between 0.50 and 0.90, and trigger human review below 0.50.
3. Replace Expensive LLM Router Prompts With Pointer Scoring: Audit your multi-agent orchestration graph for nodes that call frontier models solely to classify intent or pick one of five sub-agents, and swap those nodes to Strands Decider 2B to eliminate token generation costs.
4. Fine-Tune on Internal Audit Logs Using the Open Recipe: Use AWS’s published training scripts and LoRA recipe to adapt the 1M-parameter pointer head on your company’s historical support triage or tool-rejection logs.