How is Strands Decider different from asking an LLM to pick an option?

Strands Decider 2B cannot write text at all. AWS removed the language-model head from Qwen3.5-2B and added a pointer head of about one million parameters that scores the options you supply in a single forward pass. Every answer is one of your options plus a calibrated confidence, so there is no output to parse and no invalid answer.

How fast is Strands Decider 2B on local hardware?

The Strands Decider README reports a 115 ms median and 299 ms 95th-percentile latency per question on an Nvidia RTX 3090, and a 153 ms warm median on an M3 Pro Mac for tasks under 300 tokens. AWS says latency grows roughly linearly with the size of the task.

How does Strands Decider score on JevBench?

Strands Decider 2B (v19) gets 72.3% (167 of 231 tasks) on the public JevBench set, with a Brier score of 0.342 and an ECE of 0.052. The AWS launch post says that places it 3rd of 33 models in the 2B class, and 1st of 30 when the just-over-2B models are left out.

Can I retrain Strands Decider on my own hardware?

Yes. The strands-decider repo ships the training setup and a list of data sources. The README says a full training run takes about 11 hours on a single 24 GB RTX 3090, or about 1 hour 10 minutes on eight H100s in FAST mode. The Hugging Face card lists public datasets such as GLUE, PAWS and HotpotQA.

AWS Strands Labs · 2026-10-01 · major

Strands Decider 2B — AWS opens a local decision model for agents

Strands Decider 2B is an Apache-2.0 decision model from AWS Strands Labs. It picks between options or rates on a scale with a calibrated confidence, in about 115 ms on an RTX 3090, and runs locally.

Strands Decider launch graphic from the Strands Agents blog

A 2B open model that answers choice, yes/no and rating questions with a confidence score, fast enough to run before every agent step.

Key specs

Median latency (rtx 3090)	115 ms
Jev bench accuracy	72.3%

Quick facts

Maker	AWS Strands Labs
Base model	Qwen3.5-2B + rank-16 LoRA + pointer head
License	Apache-2.0
Question types	choice, noul (yes/no), score
Runs on	NVIDIA GPU, Apple silicon (MLX), CPU
Install	pip install strands-decider

What is it?

Strands Decider 2B is a small open-source decision model from AWS Strands Labs, released on October 1, 2026. Instead of generating text, it picks one of the answers you give it, or rates something on a scale, and returns a calibrated confidence for every answer. It answers the same kind of question as TypeSafe's hosted Jev model, but it is free to download and runs locally.

How does it work?

AWS starts from the Qwen3.5-2B base model, removes the head that predicts the next word, and adds a pointer head that compares hidden states to score each supplied option. A rank-16 LoRA adapter is trained on top. One forward pass answers the question, with no decoding loop. A Python CLI and a `/v1/systemone` HTTP server ask choice, noul (yes/no) and score questions.

Why does it matter?

Agent builders often call a full LLM just to route a request, pick a tool or check an argument. A local decision model does those small calls in about 0.1 seconds with a confidence they can act on, and sends only hard cases to a larger model. Strands Decider puts AWS next to Cloudflare's Clef, Jeff and Laya in a fast-growing group of open decision models.

Who is it for?

agent builders, ML engineers

Frequently asked questions

How is Strands Decider different from asking an LLM to pick an option?: Strands Decider 2B cannot write text at all. AWS removed the language-model head from Qwen3.5-2B and added a pointer head of about one million parameters that scores the options you supply in a single forward pass. Every answer is one of your options plus a calibrated confidence, so there is no output to parse and no invalid answer.
How fast is Strands Decider 2B on local hardware?: The Strands Decider README reports a 115 ms median and 299 ms 95th-percentile latency per question on an Nvidia RTX 3090, and a 153 ms warm median on an M3 Pro Mac for tasks under 300 tokens. AWS says latency grows roughly linearly with the size of the task.
How does Strands Decider score on JevBench?: Strands Decider 2B (v19) gets 72.3% (167 of 231 tasks) on the public JevBench set, with a Brier score of 0.342 and an ECE of 0.052. The AWS launch post says that places it 3rd of 33 models in the 2B class, and 1st of 30 when the just-over-2B models are left out.
Can I retrain Strands Decider on my own hardware?: Yes. The strands-decider repo ships the training setup and a list of data sources. The README says a full training run takes about 11 hours on a single 24 GB RTX 3090, or about 1 hour 10 minutes on eight H100s in FAST mode. The Hugging Face card lists public datasets such as GLUE, PAWS and HotpotQA.

Try it

pip install strands-decider