Strands Decider 2B: A 2B-Parameter Decision Model for Agentic Workflows

Strands Labs (AWS) released Strands Decider 2B, a 2-billion-parameter open-source decision model optimized for fast local experimentation and agentic workflows. Unlike general-purpose LLMs, decision models are designed to pick between a fixed set of options (e.g., "Is 'turn on the lights' about the coffee machine? Yes or no.") or assign simple numerical scores (e.g., sentiment between 0 and 1).

What it is and why it matters

Decision models trade flexibility for speed and reliability. They always return an answer from the provided options, run with very low latency, and emit a calibration score ("how sure can I be this is correct?") that frontier LLM APIs don't expose. The tradeoff: they're worse at complex reasoning and can't generate text, so they're unsuited for coding, chatbots, or summarization.

Strands positions this as a new class of "system one" models, following TypeSafe AI's launch of Jev earlier this month. The intent is to drive agentic decision steps in the Strands Harness SDK.

Architecture

The model takes a pre-trained Qwen3.5-2B torso, strips the LM head (removing text generation), and replaces it with a pointer head that scores each offered option. The head compares the hidden state at each option position against the hidden state at the <answer> position. The head is small — just over 1M parameters — and the torso is fine-tuned with a rank-16 LoRA adapter.

This is v19 of the architecture. The first iteration used a slot head, which performed significantly worse; all iteration notes are in the repo.

Benchmarks

Strands measured accuracy on JevBench's public set and calibration via Brier score on the same set:

• Accuracy: 3rd of 33 in the 2B class
• Accuracy excluding just-over-2B models: 1st of 30
• Latency (median): ~115ms on an Nvidia RTX 3090, ~153ms for small tasks on an M3 MacBook
• Latency scales roughly linearly with task size

Availability

• Code on GitHub: strands-decider-2b
• Weights on Hugging Face
• Includes all training data and scripts
• Runs on local CPU or GPU — no API calls required

Who it's for

Developers building agentic workflows who need fast, calibrated yes/no or classification decisions at the edge, and researchers who want a small, hackable base for decision-model experiments.