What’s new here

Strands Decider takes a pretrained 2B decoder, discards the language-modelling head, and attaches a pointer head that scores each option by comparing hidden states at the <answer> position against each option’s last token. One forward pass, no decoding loop. Every answer comes with a calibrated confidence: the README reports that answers at 0.9 or above are correct about 95% of the time on short classification tasks the model has never seen.

What it does

The model answers three question types against a text state:

• choice: pick one of N named options
• noul: a yes/no probability (0 to 1)
• score: map state to a numeric position on a labelled rubric

You can batch all three types in one call; the state is read once and each question adds only its own tokens, so many questions against one text are nearly as cheap as one.

The 2B Qwen3.5 model runs at a median 115 ms per question on an RTX 3090 and also serves on Apple Silicon via MLX or on CPU. Gemma 4 variants (E2B through 26B-A4B) are available for higher accuracy at the cost of more memory. A strands-decider serve command exposes an HTTP endpoint so you can ask questions via HTTP requests.

Who it’s for

Engineers building agentic workflows who need fast, cheap classification inside the loop: routing support tickets, gating tool calls, running guardrails, scoring LLM outputs, or triaging requests. The README positions it for decisions where you would otherwise send a short prompt to a frontier LLM just to get a yes/no or a category back.

Try it

Install and ask a question from the CLI:

pip install strands-decider

strands-decider ask StrandsAgents/strands-decider-2B-qwen3.5-v1-2610 \
  --state "Help! My payouts have been failing for 3 days! " \
  --choice "Which team should handle this?=billing,sales,retail" \
  --noul "Does this convey urgency?" \
  --score "How frustrated is the writer?=calm,frustrated,depressed"

Run as a server:

strands-decider serve StrandsAgents/strands-decider-2B-qwen3.5-v1-2610 --port 8000

Gemma 4 models are on main but not yet on PyPI:

pip install "strands-decider @ git+https://github.com/strands-labs/strands-decider"

On Apple Silicon, add --device mlx for 1.4 to 1.6x faster than MPS (requires the mlx extra, currently install-from-source only).

How mature is it

556 stars, 72 forks, 20 open issues, 17 open pull requests, 6 contributors, 22 commits in the last 90 days. Created 2026-09-29, last pushed 2026-10-09. No formal releases yet. Licensed Apache-2.0. The training record documents every run since v9 with pre-registered predictions, which gives you a clear history of what worked and what did not.