Contents
Strands Decider 2B is an open-source model with 2 billion parameters that selects answers from options instead of generating text. It runs on your own CPU or GPU, making it useful for testing low-latency decisions such as tool selection and input checks.
Key Points#
- Its 2 billion parameters make local execution and fine-tuning accessible.
- It returns scores and confidence for the options and efficiently handles multiple questions about the same input.
- On the public JevBench set, it ranks third among 33 models in the 2B class, and first among 30 when models slightly larger than 2B are excluded.
- Local measurements show a median of about 115 ms, and about 153 ms for small tasks on an M3 MacBook.
- It supports predefined decisions rather than handling chat, summarization, or complex reasoning.
What the Model Does#
Typical LLMs generate text freely. Decider instead selects one of the supplied candidates or answers binary questions. For example, it can classify an inquiry to route it to “billing,” “sales,” or “retail.”
It returns a score for each candidate alongside the answer, letting you consider both the selection and how much to trust it. Because it always answers from the provided options, it is easy to integrate into workflows with constrained output formats.
The trade-off is that it cannot write free-form text. It is unsuitable for code generation, document summarization, or conversation. It also struggles with step-by-step reasoning on difficult problems, so it is best viewed as a model that handles some of an agent’s decisions rather than a replacement for an LLM.
How the 2B Model Works#
It uses Qwen3.5-2B as its base, with the text-generating LM head removed. A small pointer head is added to score candidates. The pointer head has just over 1 million parameters, and the base model is fine-tuned using rank-16 LoRA adapters.
The public release is v19. This version is the result of successive improvements to the initial slot-head approach, and the repository records the changes in each version. The training data and scripts are also public, so you can explore how the model was built or train it yourself, as well as run inference.
Accuracy and Speed#
The published evaluation uses the public JevBench set to assess accuracy and confidence calibration. Decider 2B ranked third among 33 models in the 2B class, and first among 30 when models slightly larger than 2B were excluded. It reportedly answered 100% of the benchmark’s simpler tasks correctly, but these are benchmark results and do not guarantee the same accuracy on real business data.
Processing time increases roughly linearly with task size. Measurements on an Nvidia RTX 3090 showed a median of about 115 ms, while small tasks took about 153 ms on an M3 MacBook. Both are local measurements from the publisher, so results will vary with hardware, input length, and implementation. Measuring again in your deployment environment is the most reliable approach.
Try the CLI First#
If you have a Python environment, you can install the CLI and run a short multiple-choice task.
pip install strands-decider
strands-decider ask StrandsAgents/strands-decider-2B-hobson-v19 \
--state "3日間、支払いが失敗しています。" \
--choice "担当チームは?=billing,sales,retail"The output includes the selected candidate, confidence, and per-candidate scores. In this example, billing is selected with a confidence of 0.768, while the candidate scores are 0.845 for billing, 0.091 for retail, and 0.064 for sales. Confidence and candidate scores are displayed as separate values, so check what each means before using them in a workflow.
Potential uses include model routing, tool selection, evaluation, guardrails, and memory or context management. A setup that leaves complex decisions to an LLM and routes routine decisions to Decider may reduce latency and cost.
Considerations for Agent Integration#
The Strands sample uses Decider to check tool calls just before execution. When asked about the weather, the example agent tries to infer a location and call a tool, but Decider checks whether the arguments are grounded in the user’s message and whether it is too early to call the tool without clarification. If no location is given, it blocks the call and prompts the agent to ask the user.
The sample is a small demo with manually configured questions, thresholds, and policies. For deployment, design questions and decision criteria around the impact of incorrect decisions, and validate accuracy and confidence on real data. Rather than authorizing consequential actions solely on model scores, provide a confirmation step where needed.
Next Step#
Install the CLI, prepare a few multiple-choice tasks close to your use case, and check the answers, scores, and processing times in your own environment.
Source: https://strandsagents.com/blog/introducing-strands-decider/
Contents
