Amazon Web Services' Strands Labs has open-sourced Strands Decider 2B, a small "decision model" that selects one answer from a predefined list and reports how confident it is rather than generating text, TechCrunch reported Thursday. The release puts Amazon into a crowded new category that took off with startup TypeSafe's Jev model and that OpenAI also entered earlier in the week.

What Happened

According to TechCrunch, the project started with Amazon distinguished engineer Marc Brooker, who built his own version of a decision model after coming across Jev. That homebrew model briefly sat at the top of the Jevbench leaderboard for models of its size, and Amazon then turned it into a formal release.

The model is built on Qwen3.5-2B, the 2 billion-parameter open model from Alibaba's Qwen family. SiliconANGLE reported that AWS removed the usual language-model head, which produces tokens, and replaced it with a custom "pointer head" of about 1 million parameters. The base model was fine-tuned with a rank-16 LoRA adapter so that it scores its internal representations directly against the positions of the candidate answers.

RelatedClaude Code Desktop Preview Flaw Could Have Exposed Developers’ Cloud Keys, Researchers Say

In a post on the Strands Agents blog, Brooker and co-authors Mike Chambers and Fabio Nonato de Paula put median latency at about 115 milliseconds on an Nvidia RTX 3090 graphics card and about 153 milliseconds on an M3 MacBook for small tasks. Weights are posted on Hugging Face, and the code, training scripts and examples are on GitHub, with a pip package for installation.

Why It Matters

Many steps inside an AI agent are multiple-choice questions in disguise: which model should handle a request, which tool to call next, whether an action breaks a policy. Developers often send those questions to a large language model, paying for tokens and waiting for a full response. A decision model answers with a single choice plus a score.

Brooker told TechCrunch that the confidence score lets such a workflow step be structured more reliably, with lower latency and potentially lower cost for AWS customers running agent workflows. The Strands team says frontier LLM APIs do not offer comparable reliability scores, and it measures calibration, or how well stated confidence matches actual accuracy, with Brier scores.

AWS lists model routing, tool selection, evaluations, guardrails, memory and context management, and policy classification as early uses. It describes a "hybrid agent" pattern in which the decision model handles simple, repeated choices and an LLM takes the harder reasoning. In the Strands Agents SDK, the model can plug into the framework's intervention system and check a tool call before it runs, acting as a guardrail.

A Crowded Field

TechCrunch said dozens of comparable models have appeared since Jev's debut. TypeSafe CEO Diogo Almeida responded to the wave of big-company entrants by saying they may be underestimating how hard it is to make such models genuinely smart.

The benchmark picture is still moving. While Brooker's early version briefly led Jevbench in its class, the Strands blog says the released model ranks third of 33 entries in the 2 billion-parameter class on the benchmark's public set. SiliconANGLE reported that it scored well on accuracy and calibration against other open 2B models.

What to Watch

Because the weights are open, developers can test AWS's latency and calibration claims on their own hardware and tasks, and those results will matter more than leaderboard snapshots. Watch how Jevbench rankings shift as OpenAI's offering and other rivals are measured, and whether the hybrid design, with small decision models gating calls to larger LLMs, shows up in production agent frameworks beyond Strands.