AWS open-sources Strands Decider 2B, a small local decision model

The Strands Agents team at AWS released Strands Decider 2B on October 1, a small open-source decision model for fast experimentation and local development. It follows TypeSafe AI's launch of Jev earlier this month, which started interest in this class of models.

What a decision model is

Unlike an LLM that generates free text, a decision model picks between a given set of options and assigns numerical scores. Examples from the announcement: is this text about the coffee machine, yes or no; which language is this phrase in; is this sentiment positive on a scale from 0 to 1. Each decision comes with a reliability score that frontier LLM APIs do not offer. The trade-off is clear: it is faster and sharper at a given size, but weaker at complex reasoning and unable to write text, so it is not suited to coding, chatbots or summaries.

Key facts

• Size and design: 2 billion parameters. It starts from a Qwen3.5-2B torso, removes the language-model head and adds a small pointer head of just over a million parameters that scores each option. Fine-tuning uses a rank-16 LoRA adapter.
• Speed: a median of around 115 ms for small tasks on an Nvidia RTX 3090, and about 153 ms on an M3 MacBook.
• Quality: third of 33 models in the 2B class on JevBench's public set for accuracy and calibration, and first of 30 when models just over 2B are excluded.
• Openness: code on GitHub, weights on Hugging Face, and all training data and scripts. Install with `pip install strands-decider`.

How it is meant to be used

The team sees early use in model routing, tool selection, evaluations, guardrails, memory and policy classification. Their example sits in front of a tool call: before a weather tool runs, the model checks whether the arguments came from the user and whether it is too early to call, and the agent asks which city instead of guessing.

Why it matters

Agents spend a lot of time on small yes/no choices. Using a giant model for each is slow and costly. A cheap local judge that can say "I am 77% sure" is a practical building block for hybrid agents.

Dany's take

This is the unglamorous but useful part of the agent boom. The article itself says the example is an illustration, with hand-picked questions and thresholds, so test it on your own data. I like that everything, including training data, is public. Together with Cloudflare's Clef, it shows the decision-model idea is spreading fast.

Source: Strands Agents, Introducing Strands Decider 2B