Skip to content

Open models

AWS’s Strands Decider 2B picks from your list, never writes

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: aiaffairs.com

AWS on October 1 released Strands Decider 2B, a small model that never writes a sentence. It takes a list of options supplied by the developer and scores each one, or returns a number. The decision models that came before it are hosted services. Decider is something you download, and AWS published the training data and scripts along with it.

As The New Stack describes it, the release is AWS’s entry in a wave that TypeSafe’s Jev started a few weeks ago, and that Kev, imajev, Laya and others have joined since. Jev went into limited early access on September 15 and became a public hosted API on September 21. It is proprietary. OpenAI’s Decisions API, launched on September 29 as a limited preview, is also hosted, and it points OpenAI’s Luna model at questions with predefined answers.

All of these models give up free-form text in exchange for one of two outputs: a choice among the options the developer listed, or a numerical score. That makes them good at the small, frequent calls inside an agent: routing a natural-language request, picking a tool, grading an output, checking a proposed action. Conversation and longer work stay with a generative model.

AWS’s version starts from Qwen3.5-2B, which the team calls the “torso”, meaning the part that understands language. They removed the language-model head, the final layer that produces text, and put a pointer head in its place. The pointer head scores the options it is given and has just over a million parameters. The backbone itself was tuned with a rank-16 LoRA adapter, a small set of trainable low-rank weights added on top of a model whose original weights stay frozen.

TypeSafe says it reached the same goal through a different design. Its launch post describes a parallel sampler that “Generates all outputs in a single query” instead of producing tokens one at a time. Jev also accepts typed questions in three forms: true/false, an unordered choice, and an ordered score.

A closed answer list means Decider cannot make up an option nobody supplied. It can still pick the wrong one. The New Stack calls that a minor tradeoff, since LLMs aren’t always right either, and points to what developers get for it: faster decisions, plus confidence scores their code can act on. AWS says it aimed for a balance of accuracy, calibration and latency. Calibration is how closely the model’s confidence matches how often it is actually right.

AWS built its demo on its open-source Strands agent framework. A user asks for the weather and doesn’t say where. The agent guesses a city and proposes a call to a weather tool. Before the tool runs, Decider checks two things: whether the argument values come from the conversation, and whether the agent has enough information to go ahead. The application then sends the agent back to ask the user which city they meant.

The check runs through Strands’ intervention system, which gives the developer four options for a proposed tool call: let it proceed, deny it, ask a human to confirm it, or send feedback to the agent. In the demo, Decider runs locally and the agent reaches its generative model through Amazon Bedrock. AWS says it is also working on integration libraries for decision models.

The performance figures come from AWS. On an Nvidia RTX 3090, Decider makes a decision in under 100 milliseconds, and response times grow as tasks get bigger. On an M3 MacBook, the median for small tasks is around 150 milliseconds. In September, OpenAI said its Decisions API picks an answer in 150 milliseconds, from its own servers. TypeSafe gives Jev an end-to-end response time of 70 to 500 milliseconds.

On the public set of JevBench, AWS says Decider ranks second among public models of roughly 2 billion parameters. It ranks first among public models that come with a full training recipe. AWS also says Decider answers every question in JevBench’s easy tier correctly, and that tier is closest to the routine agent decisions the model was built for.

The base model matters here. Kev, like Decider, is built on an open Qwen model, which shows how much of this experimentation now depends on open weights. Kev already supports local deployment and fine-tuning. AWS goes a step further by publishing the data and scripts, so developers can inspect how the model was trained and adapt the recipe to their own decisions.

Decider came out of Strands Labs, the unit AWS opened earlier this year for experimental approaches to agentic AI. It also follows AWS’s recent release of Strands Harness, which bundles the tools and supporting machinery for running longer-lived agents.

Hosting is where the release is still thin. The demo’s split, with Decider on localhost and the generative model in Bedrock, works for experiments. But a team taking an app built on Decider to production will want a hosted version, and AWS hasn’t said whether this model, or a later one, will be offered in its cloud.

The model released now is the second major iteration of the architecture. AWS says an earlier head design performed significantly worse, and it left every earlier iteration in the repository so developers can trace how the model evolved.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.