System One new

What it is

A System One model answers typed questions about a piece of state and returns probabilities. It never writes text. You declare the possible answers up front, so there is nothing to parse and nothing to repair.

TypeSafe AI coined the name on 15 September 2026 when it shipped Jev, after Kahneman’s System 1: the fast, automatic judgment that does not deliberate. The category is the other half of an agent stack whose frontier models got slower and more expensive as they got better at reasoning. Most decisions inside an agent loop — is this tool call risky, which model does this request deserve, is this ticket a refund — do not need deliberation. They need an answer in 30 ms.

Within a week of Jev’s launch a dozen open models copied the idea. The category is now real; a shared spec is not.

How it works

Every model in the category has the same shape:

• State in. Text or JSON. One encode pass.
• Typed questions. Three primitives everyone converged on: Choice (pick one of a described set), Score (place it on an ordered rubric), Noul (is this statement true).
• Probabilities out. The selected option, a probability for every option, and a confidence number derived from how concentrated the distribution is.

Questions are evaluated in parallel against one encoding of the state, so asking five questions costs about what one costs. Architecturally the open models are almost all a bidirectional encoder — usually ModernBERT-large — with a decision head that scores each option at its own mask position. Options are supplied at request time, which is what separates these from a fine-tuned classifier with frozen labels.

The nearest thing to a standard is TypeSafe’s wire format, POST /v1/systemone. Several projects implement it deliberately so you can repoint a base URL.

The models today

Jev — TypeSafe AI · closed API, early access · jev-1.13.0

The original. 70–500 ms, $0.042 per million input tokens with output free, 64k context. Trained with Reinforcement Learning for Calibrated Decisions on entirely synthetic data. Waitlist only; in practice people reach it through the Vercel AI Gateway (typesafe-ai/jev, via the AI SDK’s experimental_evaluate) or LangChain’s langchain-typesafe. Full page: Jev.

Laya — Convai Innovations · Apache-2.0 · ModernBERT-large

421M English, 322M multilingual, plus a typed-decisions checkpoint fine-tuned on the task. 32.8–39.5 ms per question on a T4, 193–464 ms on CPU, 100+ languages claimed and 45 of 51 tested. Expected calibration error 0.081 after temperature scaling. Convai reports it ahead of Jev on their own set (0.766 vs 0.727 typed decisions, 0.950 vs 0.910 on AG News). The honest caveat is in their own model card: the base checkpoints score near chance zero-shot (0.362) — it is a fast base to specialise, not a drop-in decision engine.

von — Victor Panisa · Apache-2.0 · ModernBERT-large 395M

Positioned as a local drop-in for Jev. Sub-15 ms on accelerated hardware, ~18 ms on Apple MPS, ~480 ms on CPU. API is decide() / judge() / rate() plus a system_one() that batches independent questions. Trained on ~290k synthetic examples across workflow, security, safety and triage. Self-reported 93.5% against Jev’s 97.2% on the same set — slower on accuracy, far faster and free.

The reproduction shelf — everything else, a week old

• openjev-sglang (ekzhang, MIT): a 35B model behind the TypeSafe API, served on a B200 via Modal. Wire-compatible, not cheap.
• NanoJev (TianyuCodings, MIT): 600M trained from scratch, weights and dataset published.
• OpenDecision (deepanwadhwa, Apache-2.0): ~400M ModernBERT, NLI-based, runs on CPU.
• SemIf (MIT, 4B) and mini-jev (MIT, 4B, Apple Silicon): logit readers over a frozen LLM — choice-type questions only, no training.
• MiniSystemOne: ~27M trained from scratch in under five hours on one 8GB GPU. Not for production; the clearest way to understand the architecture.
• Variants: jev-visual for images, jevmlx and PocketJev for Apple Silicon and iPhone.

Nobody has reproduced RLCD. It has never been described in enough detail to copy, so the open models either disclaim calibration or bolt on temperature scaling afterwards.

When to use one

Reach for a System One model when the decision is bounded, high-volume and inside a loop:

• Gating tool calls before they run, with low-confidence ones routed to a human.
• Routing: deciding which LLM a request deserves before paying for one.
• Watching an agent trace step by step — looping, drifting, off-task.
• Bulk classification: tickets, email, moderation.

Use an LLM with structured output instead when the answer set is not known in advance, when you need free text extracted, or when the decision needs several steps of reasoning. And before reaching for either, check whether a fine-tuned encoder on your own labels already wins — for one fixed task with plenty of labelled data, it usually does. The category earns its keep when the questions change at request time.

Watch out for

• Every number on this page is self-reported. Vendors benchmark against their own sets; Convai and TypeSafe both claim the lead. Run your own eval set before believing any ranking.
• Wire compatibility is not behavioural compatibility. A model that speaks /v1/systemone will accept your requests and return different probabilities. Thresholds do not transfer between models, or between versions of one model. Pin the version, log it, re-tune after a swap.
• Calibration is the whole value proposition, and the hardest part to copy. If you are going to act on confidence > 0.9, bucket answers by confidence on labelled data and compare with actual accuracy first.
• Open weights mean you own the fine-tuning. Most of these need task data to beat a prompt; the base checkpoints are near chance on general typed decisions.
• “Cannot hallucinate” only means it cannot invent an option. It can still choose the wrong one, confidently.
• State is not adversarial input. A prompt injection in user-controlled state steers the decision. Never make this the only guard in front of a dangerous action.
• No arithmetic, and literal reading. Counting, dates and negation are unreliable across the whole category. Do them in code.
• The shelf is a week old. Most of these projects have one author and no users. Treat anything outside Jev and Laya as a starting point you will maintain yourself.