Amazon releases its own Jev clone as decision models flood the web

TechCrunch reports Amazon Web Services released Strands Decider 2B from Strands Labs the same week OpenAI announced a similar class of model. Built on a Qwen3.5-2B “torso,” it outputs calibrated choices among pre-set options with confidence scores — open source, local-runnable, aimed at agent workflow steps that don’t need a full frontier LLM.

Distinguished engineer Marc Brooker started it after seeing TypeSafe’s Jev; an early homebrew briefly topped Jevbench in its size class. TypeSafe CEO Diogo Almeida tells TechCrunch he still doesn’t see real competition — many clones look like cool architectures, not dedicated intelligence work.

Decision models are a healthy specialization. Not every agent step should burn a frontier call when the job is “given these tools, pick the next one.” We’re for open, local deciders with confidence scores if they actually reduce wrong branches in production workflows.

The flood of Jev-likes is both validation and a warning. Cheap to train doesn’t mean calibrated or multilingual-smart, as Brooker notes — push accuracy too hard and you gut the general understanding that makes the torso useful. Almeida’s dig that clones underestimate difficulty sounds like founder defense and may still be right.

I’m watching Jevbench and real AWS customer designs: do Strands Decider steps cut latency and cost without silent misroutes? Useful AI often looks like a boring 2B model that knows when it’s unsure.

TypeSafe named Jev after economist William Stanley Jevons; the broader thesis is cheaper intelligence increasing demand for decisioning inside agents.