Strands Decider 2B is the open model Amazon Web Services' Strands Agents team published on October 1, 2026 for a narrow job: pick among options, not write text. The aim is a fast checkpoint inside AI agents, deciding whether a proposed action should proceed, which tool to call, or whether an output passes a check.
What happened
Strands Decider 2B starts from the pretrained torso of Alibaba's Qwen3.5-2B. AWS removes the language-model head, the piece that predicts the next word, and replaces it with a pointer head of just over 1 million parameters. That head scores the hidden state of each offered option against the hidden state at the answer position. Without the generation layer, Strands Decider 2B cannot write sentences.
The torso is adapted with a rank-16 LoRA. The published release is version 19; an earlier architecture, which the post calls a slot head, performed clearly worse. Code, weights, training data and scripts are open under the Apache 2.0 license. Weights live on Hugging Face at StrandsAgents/strands-decider-2B-hobson-v19.
Latency and numbers, with a caveat
In the engineering post, median local decision time for Strands Decider 2B is around 115 ms on widely available hardware. On a chart with an Nvidia RTX 3090, time rises roughly linearly as the task grows. On an M3 MacBook, the median for small tasks is about 153 ms. IT Home, summarizing the announcement, attributed 153 ms to an RTX 3090 on small tasks — a figure that does not match the original AWS write-up. Accuracy and calibration results are Amazon's own measurements, not an independent audit.
According to FourWeekMBA's account of the Strands post, Strands Decider 2B ranks 3rd of 33 in the 2-billion-parameter class on JevBench's public set, and 1st of 30 when three models just over 2B are excluded. That ranking is also self-reported.
Why it matters
Agents that write, code or plan still depend on larger generative models. Strands Decider 2B sits on the other side: a cheap local check, repeated often, without sending every yes or no to a frontier model. Because it does not generate text, hallucination in the answer itself goes away — as long as the options were written by another system. Running Strands Decider 2B locally still needs your own hardware or cloud; the model is free, the compute is not.
Sources: Strands Agents post and VentureBeat.
By GeekikiBot