JEV: an open decision model after closed Jev opened the door

On 15 September 2026, TypeSafe AI opened a new door with Jev: a model that does not write a paragraph, but returns a structured decision with a probability. Less than three weeks later, Hugging Face listed an open-weights model from the JEV family, named GEV-26B-Decide, from AutoTrust AI — roughly the same decision class, except you can run it yourself.

This note is not a marketing bake-off. It records what we can verify today from official cards and announcements, and it separates confirmed facts from company claims or uncertain items.

What does JEV do?

JEV-26B-Decide is a decision model, not a chat assistant.

You give it a state (text, an image, or a mix) and a typed question; it returns a calibrated probability for every option — not open-ended prose you have to parse.

Question kinds documented on the model card:

• yes / no (noul)
• choice among 2–256 options
• score on a 0–5 scale

It works over text and images. The card states that the decision head was trained on text, and that image decisions are zero-shot — a company-stated capability, not an independent external audit.

System 1, then thinking when needed

As AutoTrust describes the design:

1. System 1: one forward pass — the card cites about 45 ms on a B200-class GPU (company figure on specific hardware).
2. Adaptive thinking (optional): if the leading option’s confidence is below 0.8, System 2 runs (the same backbone in Gemma 4 thinking mode), then mixes the two distributions equally.
3. System 2 alone: the unmodified google/gemma-4-26B-A4B-it for text and reasoning.

For builders: classify and route fast; pay for thinking only when the model is unsure.

Origin, naming, dates

• Base: Google Gemma 4, about 26B parameters, roughly 4B active per token (MoE as described on the card).
• First published as autotrust/JEV-Gemma4-26B-A4B.
• On 3 October 2026 renamed to autotrust/GEV-26B-Decide — same weights, per the card.
• The same day, the card added computer-use and robot-arm demos (simulation), with claimed System 1 step times around 60–85 ms.

Important: AutoTrust states explicitly that JEV is independent of TypeSafe AI — not affiliated, endorsed, or a TypeSafe product. The model belongs to AutoTrust’s JEV family; its Hugging Face page currently carries the name GEV-26B-Decide (first published as JEV-Gemma4-26B-A4B). The similar name does not make it a clone of closed Jev.

Uncertain from the card: a named individual founder/author for AutoTrust; the release is under the organization.

Closed Jev opened the door — what’s the contrast?

	TypeSafe Jev	AutoTrust JEV-26B-Decide
Public launch	15 September 2026 (company announcement)	3 October 2026 under the name GEV-26B-Decide on Hugging Face (after the earlier JEV-Gemma4-26B-A4B publish)
Form	Hosted / closed weights (API)	Open decision weights (adapter + head)
Class	TypeSafe’s first “System One”	Open decision model in the same wave
Outputs	Typed decisions with probabilities (choice / score / noul)	Same general idea, with a local /v1/decide path
Claimed speed	70–500 ms end-to-end (TypeSafe claim)	~45 ms System 1 on B200 (AutoTrust claim)
Claimed training	RLCD (Reinforcement Learning for Calibrated Decisions)	Decision head + LoRA on a frozen Gemma backbone (details on the card)
Vision	Not the focus of the first launch	Text + images on one engine

TypeSafe context (for orientation only)

• Left stealth in September 2026 with a $40M seed led by DCVC (per TypeSafe pages).
• Founder Diogo Almeida, former OpenAI researcher, cited as a contributor to RLHF work linked to ChatGPT (company biography).
• Jev is meant to be called by software: state + questions → structured answers, not chat.

Is JEV “the first of its kind”? No. TypeSafe announced the System One class and Jev first. JEV is an open alternative you can host, with adaptive thinking and vision on Gemma 4 — newsworthy without “world’s first” hype.

What you need to run it (per the card)

• Decision adapter, head, and calibration files: Apache-2.0.
• Base model: Gemma 4 terms.
• Documented server path typically needs a GPU with about 80 GB+ memory and a patched vLLM build for Gemma 4.
• Native context 256K tokens; the card tests decisions up to 128K and says longer contexts are untested.

These are engineering requirements, not end-user product promises.

Why it matters for builders

The September–October 2026 wave says something simple:

Not every intelligence needs a paragraph.
Much of automation is fast decisions in the loop: route a ticket, call a tool?, which UI element to click?, raise confidence before acting?

Jev proved commercial demand via API.
JEV says: you can own the decision layer locally and wire it to screens and robots — if you bring the hardware and ops.

For products and agents: put cheap fast decisions in System 1; keep generative models for writing and heavy reasoning. Don’t replace one with the other; stack them.

Useful links

• Model card: autotrust/GEV-26B-Decide on Hugging Face
• Catalog note: GEV-26B-Decide on ModelSystem.One
• Short launch note: Jev News — GEV-26B-Decide
• What Jev is (TypeSafe): What is Jev?
• TypeSafe announcement: Introducing System One Models & Jev
• Base license: Gemma 4 license

On images and video

The Hugging Face card includes demo videos (adaptive thinking, computer use, robot arm). Media files are not copied here; watch them on the model page so we don’t redistribute assets without a clear license to do so.

Bottom line

Closed Jev opened the door.
Open JEV lets you walk in and run it yourself.

It is still early to judge quality beyond company numbers and their boards. What is worth watching: whether a calibrated decision layer sticks in real products — routing, guardrails, agents, visual control — or stays a benchmark wave.

I will update this note when independent, verifiable runs appear, or when we test the model ourselves and log the result in the build log.

Cover image: official image from the model's page on ModelSystem.One.

---

Abdulrahman Alshehri — Building What's Next.