Amazon Web Services introduced Strands Decider 2B, a small open-source decision model, in its weekly roundup on Oct. 5. AWS says it’s optimized for fast experimentation and local development. The model’s own launch post on the Strands Agents blog is dated Oct. 1.
The 2B refers to about 2 billion parameters, the internal values a model learns in training. That’s small enough to run on a single computer’s CPU or GPU, according to the post.
What a decision model is
AWS describes Strands Decider as one of a new class of “decision models or system one models.” The term borrows from psychology, where system one means fast, automatic judgment. The post says the category has drawn attention since TypeSafe AI launched a model called Jev earlier in the month.
Unlike a chatbot model, a decision model doesn’t write text. It picks between set options or gives a score. The post’s examples include deciding whether “turn on the lights” is about a coffee machine, or naming the language of a phrase from a short list.
The Strands team says that trade-off makes such models faster and more capable for their size, and they always return one of the offered answers. It also makes them “significantly worse at solving complex problems than reasoning models” and unsuited to coding or summarizing documents, the post says.
Each answer comes with a confidence score. The GitHub project says that on short classification tasks the model hadn’t seen, answers given with a confidence of 0.9 or higher were right about 95% of the time. It suggests checking with a person below that level.
How it’s built and how it scores
The model starts from Qwen3.5-2B, an open model from Alibaba. The team removed the part that generates text and replaced it with a small “pointer head” of just over a million parameters that scores each option.
It then fine-tuned the rest with LoRA, a method that trains a small add-on instead of the whole model. The released version is v19.
On JevBench’s public test set, the team says the model placed third of 33 in its size class for accuracy and calibration, and first of 30 when slightly larger models are excluded. Calibration measures how well a model’s confidence matches how often it’s right.
The post says the model makes local decisions in a median of about 115 milliseconds on an Nvidia RTX 3090 graphics card, and about 153 milliseconds for small tasks on an M3 MacBook. Time rises roughly in line with the size of the task.
AWS released the code on GitHub and the weights on Hugging Face, along with the training data and scripts. The project can be installed as a command-line tool or run as a local server that answers HTTP requests.
The team says it’s meant for the many small choices inside agent workflows built with its Strands Agents SDK, such as routing a support ticket to the right team.
Source: AWS News Blog, Channy Yun, “AWS Weekly Roundup (October 5, 2026)”
Additional sources: Strands Agents blog, Strands Agents team, “Introducing Strands Decider 2B: a small, open source, decision model”; GitHub, strands-labs, “strands-labs/strands-decider”
