Amazon Web Services has open-sourced Strands Decider 2B, a two-billion-parameter “decision model” that answers multiple-choice, true/false and scoring questions with calibrated confidence numbers — and never generates text. It launched on the AWS Strands Agents blog (2 October NZT) with weights on Hugging Face and the full codebase, training data and training scripts on GitHub, aimed at running on a local CPU or GPU with decisions returning in tens of milliseconds. TechCrunch’s report is blunt about the lineage: it is an open-source take on Jev — the decision model TypeSafe AI unveiled just three weeks ago — which has already spawned enough imitators that TechCrunch describes the category as “flooding the web.”
🔍 THE BOTTOM LINE: A model category TypeSafe invented three weeks ago now has Amazon shipping weights for free and OpenAI selling an API preview — the “clone wars” lasted about ten days.
What it is and how it was built
Decision models — TypeSafe’s term is “System One models” — keep an LLM’s text understanding but swap the text generator for a pointer head that scores a fixed set of options. The trade, as the AWS post lays out: no coding, no chat, no summarising, and much worse at complex reasoning than a full LLM — but always an answer from the chosen options, low latency, a reliability score on every decision, and near-free parallel questions. AWS built it by taking a pre-trained Qwen3.5-2B “torso,” deleting the LM head, and bolting on a pointer head of roughly one million parameters, fine-tuned with a rank-16 LoRA adapter. The released model is v19 of a fast iteration loop whose full changelog ships in the repo. AWS claims third of 33 in the 2B class on JevBench accuracy-and-calibration (first among strictly-2B models), 100% on the benchmark’s easy tier, and median local latency around 115ms on an RTX 3090 (about 153ms on an M3 MacBook).
The origin story is the telling part. AWS distinguished engineer Marc Brooker built a homebrew decision model after seeing Jev, documented the attempt on his own blog, watched it briefly top the Jevbench ranking for its size class, and then had it cleaned up and released under Strands Labs — AWS’s experimental agentic-AI arm. “What originally piqued my interest in this class of models was that they make a perfect decider for a workflow step — ‘what is the next thing for me to do here, based on where I am?’” Brooker told TechCrunch. AWS’s own framing with customers: agentic workflows do not need the capability or cost of a full LLM at every step — routing, tool selection, guardrails, memory and context management are all choice problems.
The context: a three-week-old category with a giant on each side
The timeline is compressed past the point of parody. TypeSafe AI exited stealth on 15-16 September (NZT) with US$40 million and shipped Jev within days; Simon Willison’s write-up — which argues “decision models” is the better name — noted Jev charges only for input (US$0.042 per million tokens, cheaper than GPT-5 Nano’s $0.05) with output free. OpenAI answered with a limited-preview “Decisions API” in a Sam Altman aside at DevDay on 30 September, described by TechCrunch as a Jev-style classifier over a predefined option set, aimed at its Luna model’s agent decisions. And now Amazon — the third major name into a niche that did not exist at the start of the month. The security angle is already live: a hackathon demo last weekend used Jev to check every agentic action for under $3 per run where frontier-LLM monitoring cost roughly $372, per TechCrunch — decision models as cheap always-on copilots watching the expensive agents. That use case matters more after the July sandbox-escape incidents and the FTC probe into the labs: if every agent action needs auditing, the auditor has to cost almost nothing. This site has followed the decision-model economics from the start — OpenAI codifying a new AI role to operate its agent fleets and coding agents costing more than the developers they replace — and both problems get cheaper the moment the routine judgments between big-model calls cost a fraction of a cent.
For New Zealand the relevant angle is local-run economics: a 2B calibrated classifier that runs on a laptop CPU is exactly the shape of tooling that can sit inside an NZ health agency’s or insurer’s on-prem workflow — triage, claims routing, document classification — without any data leaving the building. The counterweight is the one Willison flags: a bare confidence number hides its reasoning entirely, so bias audits get harder, not easier, exactly as these models start making back-office decisions about people.
Our take: two things are worth more than the press release. First, the price signal — TypeSafe called its product decision models; Amazon proved that decision models are a commodity: a good-recipe 2B pointer head on top of an open-weights torso (Qwen3.5) plus training scripts is a weekend project for a strong team, which compresses margins toward zero for anyone selling “fast cheap classification” as the whole product. TypeSafe CEO Diogo Almeida told TechCrunch he does not see real competition yet — “the current batch seems more like ML people wanting to implement a cool architecture than a team deeply dedicated to making intelligence useful” — and AWS shipping a v19 with public iteration history is both that critique’s best rebuttal and its confirmation. Second, the architecture bet: the lab behaviour says LLM pipelines are re-organising into hybrid stacks — a big model for the hard steps, a tiny non-reasoning model stamping decisions on everything in between — and Amazon just handed the small tier away for free. The interesting fork is that AWS open-sourced the exact capability that OpenAI previewed as a paid API on the same news cycle; watch which model of “decision intelligence” the market actually prices.
❓ FAQ
What is a decision model? A model that takes text in but outputs only structured choices — an answer selected from options you provide, with a confidence score — instead of generating text. Faster and cheaper than an LLM for classification-style steps, and unsuited to generation tasks.
Is Strands Decider 2B really free? Yes — open-source, weights on Hugging Face, training data and scripts on GitHub, sized to run on a local CPU or GPU with ~115ms median decisions on consumer hardware.
Did Amazon copy Jev? The lineage is explicit: TechCrunch calls it a Jev clone and AWS’s blog credits Jev with kick-starting the category. AWS argues its architecture (pointer head scoring hidden states, LoRA-tuned) fixes structural shortcomings it sees in Jev’s parallel-output design.
🔍 THE BOTTOM LINE
Three weeks after TypeSafe named a category, Amazon has commoditised it — free weights, public training scripts, and a benchmark fight it’s winning on calibration. The frontier model still gets the headlines; the boring 2B model deciding what happens next in a million agent workflows may be the bigger business.
📰 Sources
- AWS Strands Agents blog (2 October 2026 NZT) — Introducing Strands Decider 2B: a small, open source, decision model
- TechCrunch (2 October 2026 NZT) — Amazon releases its own Jev clone as decision models flood the web
- SiliconANGLE (2 October 2026 NZT) — AWS debuts Strands Decider 2B, a first lightweight decision model for agentic workflows
- Simon Willison (22 September 2026 NZT) — Jev introduces a new shape of LLM — System One, aka decision models
- TechCrunch (1 October 2026 NZT) — OpenAI’s Jev clone could help the frontier lab stop its swarming agents
- TypeSafe AI blog — Introducing System One models and Jev
