Amazon releases Strands Decider, an open decision model
Amazon has published Strands Decider 2B, a small model that picks from options it is given, in the vein of TypeSafe's Jev. It is the fourth company to release a version of the idea in about two weeks.
Summary
On October 1, Amazon released Strands Decider 2B, an open-source decision model for AI agents. It takes a state (for example, a customer support message) and a list of fixed options, and returns one of them with a confidence value. It does not write free text, so it cannot invent an option that is not on the list.
Startup Fortune describes it as the fourth release of this kind in about 17 days, after TypeSafe's Jev, a Decisions API that OpenAI reportedly mentioned at Dev Day, and two Cloudflare models, Clef and Clef-flash. TechCrunch, as cited by the same outlet, called it a Jev clone.
In practice
Strands Decider answers three kinds of question: choose one option from several, answer yes or no, and rate on a scale. The project's own example sends "my payouts have been failing for three days" and asks which team should handle it. The model answers "billing", with a confidence of 0.77.
It has 1.9 billion parameters. It runs on an RTX 3090 graphics card, on an Apple-silicon Mac or on CPU alone. On an RTX 3090 the median is 115 milliseconds per question. Asking several questions about the same text costs little, because the text is read only once.
The architecture explains the difference from a language model. Amazon started from a two-billion-parameter Qwen model, removed the part that generates text and put in its place a selection head of about a million parameters that scores each option directly. One pass produces the answer, with no generation loop.
The suggested uses are the decisions inside an agent workflow: choosing which model or tool to use next, checking a call's arguments before it runs, routing requests to the right team, and classifying safety or compliance.
What we still don't know
On the public JevBench tests, the model got 167 of 231 tasks right (72.3%). It answered every easy task, 87.5% of the standard ones and 50.5% of the hard ones. The project itself warns that 231 tasks are few and that differences under about ten tasks between two runs do not support conclusions.
The confidence value is useful with care. On short questions the model has not seen, answers with confidence of 0.9 or above are right about 95% of the time, according to the project's evaluation. Below that, the README recommends confirming or asking a person.
Cloudflare's claims about Clef's latency against Jev come from the company's own tests. OpenAI's Decisions API was mentioned almost in passing by its chief executive, according to Startup Fortune, and was not confirmed in a primary source.
Why it matters
- Four companies released the same idea within weeks, which suggests that the small, fast decisions agents make have become a layer of their own, rather than a job for a full language model.
- Open models that run on ordinary hardware make it possible to try this pattern without paying per call.
- The confidence attached to each answer helps decide when software proceeds alone and when it hands the decision to a person.