strands-decider-2B-qwen3.5-v1-2610

strands-decider-2B-qwen3.5-v1-2610

Makes fast decisions for agent workflows by picking among options and rating things on a scale, with a calibrated confidence for every answer. For developers who route models, choose tools and triage requests.

Made by StrandsAgents on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Strands decider is a compact decision model that suits agent workflows and answers quickly. It does not write free text. It answers typed questions about a state by picking among options or rating on a scale, and each answer comes with a calibrated confidence. The files are a LoRA adapter for Qwen3.5-2B-Base together with a light readout head.

Who it suits: Developers who build agent workflows and want quick, scored decisions without calling a large language model each time.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Route requests and choose tools | Decide which team should handle a message
Support | Operations teams | Triage incoming messages | Flag a message that sounds urgent
AI safety | Platform teams | Add guardrails and checks | Check arguments before a tool runs

If you need a model that writes long answers, you can skip it, since it only decides.

Where to use it:

• Your own computer with Python, on CPU, CUDA or Apple MPS
• A local HTTP server started by the strands-decider command, for experiments
• Agent workflows that need routing, tool selection and evaluation

Use case | Who | What they get back
Routing | Agent developers | The best option with a confidence score
Urgency checks | Support teams | A yes or no probability
Rating | Analysts | A level on an ordered scale with probabilities

How to set it up:

• Python with pip
• The strands-decider package
• A device, where CPU, cuda or mps all work, and the base weights download at first use

1. Install the strands-decider package.
2. Ask a question with the ask command, giving it a state and one or more questions.
3. Optionally start the local server with the serve command.

pip install strands-decider

To fetch only the model files, use the Hugging Face steps.

pip install -U huggingface_hub
hf download StrandsAgents/strands-decider-2B-qwen3.5-v1-2610

How to use it: You give the ask command a state and typed questions. Pass a device flag, or let it choose the best available one. The card's example asks which team should handle a message, whether it is urgent and how frustrated the writer is.

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"

noul_0 noul = 0.858
choice_0 -> billing (confidence 0.659)
  billing                  0.773
  sales                    0.121
  retail                   0.106
score_0 score = 1.03 (confidence 0.509)
  0: calm                                     0.196
  1: frustrated                               0.574
  2: depressed                                0.230

To serve it over HTTP on your own machine, the card gives this command.

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

Tool | What it does
noul | A yes or no question
choice | Picks one of several options
score | Rates a level on an ordered scale
ask | Answers questions about a state from the command line
serve | Starts a local server for the model

Example: You pass the message about failing payouts and ask which team should handle it. The card's output picks billing with a confidence of 0.659, and rates the message as frustrated.

For newcomers and experts: If you are new, copy the ask command and swap in your own message and options. Read the confidence next to each answer to decide when to trust it.

For experts, v1 is a soup that averages three training runs. Its data mix has 246,678 rows. It learns from the output distributions of gemma-4-31B-it where they agree with the gold label, and stays anchored to the earlier v19 release through a KL term.

You give a state and typed questions
The LoRA adapter reads them
The head scores each option
You get answers with calibrated confidence

Good to know: The card reports 180 of 231 tasks right on the public JevBench. It also lists accuracy on its own held-out sets.

Title: Accuracy on held-out sets from the card
musique: 0.882 accuracy
contractnli: 0.878 accuracy
hotpotqa: 0.832 accuracy
boardgame: 0.810 accuracy
short tasks: 0.653 accuracy

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key or account is named. A CPU is enough to try it.
• Data it reaches: the base weights are fetched from Hugging Face the first time you run it. The optional server listens only on your own computer.
• Not for: exposed servers. The card says the server has no authentication, so use it only for local experiments.

On Hugging Face, 13 November 2026

• Task: text classification
• Library: peft
• Formats: Safetensors
• Downloads in the last 30 days: 0
• Likes: 18
• License: apache-2.0
• Based on: Qwen/Qwen3.5-2B-Base
• Quantized versions: 1
• Tags: peft, safetensors, strands-decider, decision-model, lora, qwen3.5, classification, calibration, typed-decisions, code, text-classification, model-index

Links

• Model: https://huggingface.co/StrandsAgents/strands-decider-2B-qwen3.5-v1-2610
• Files: https://huggingface.co/StrandsAgents/strands-decider-2B-qwen3.5-v1-2610/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download StrandsAgents/strands-decider-2B-qwen3.5-v1-2610

Then follow the usage in its model card. Nothing in this kit runs by itself.