What is Jev? A System One model that decides instead of writing

Consider classifying an inquiry that says this month's plan seems to have been charged twice. Define billing, access, and security as department choices, then enter the inquiry. Route a returned billing value to the billing queue and send low-confidence cases for human review.

A classification step differs from explaining a refund or asking follow-up questions to resolve ambiguity. If the next program only needs the department, it need not reinterpret prose; completing the support interaction still requires explanation and dialogue.

Even in JSON mode, a general LLM generates braces, keys, and values token² by token. Jev defines output structure and choices in advance and can compute multiple decisions in parallel, changing what must be generated. TypeSafe describes end-to-end latency of 70–500 ms for its service, but that is the provider's reported range, not a guarantee across regions, input lengths, or question counts.

For a comparison, use the same inquiry and number of choices and include network wait consistently on both sides. Comparing a general LLM's JSON-generation time with Jev's end-to-end time uses different measurement scopes. This guide interprets the published range; it adds no latency measurements or universal ranking.

A JEV-style decision model receives a state, question, and allowed choices, then returns a score or probability for each choice. It fits repeated ticket routing or status decisions; use a separate generator when free-form explanation is required.

The published RTX PRO 6000 single-question medians are Julia-1 3 ms, Laya 5 ms, Kev-4B 12 ms, lev 36 ms, and OpenJev 43 ms. Keep each model name with its size and language in the table, separate from service latency at other input lengths or concurrency.

Fixed choices can reduce malformed JSON or values outside billing, access, and security. But if the inquiry omits whether both charges completed, one is pending, or account access is abnormal, a billing label can still be wrong. Format validity and business correctness are separate.

High confidence still does not guarantee correctness. Route low-confidence or closely competing choices to a person, and do not automate high-impact permission or payment approvals. Have people label test cases first, compare model results to the reference, and track misclassification separately from human review.

Structured values pass readily to the next step in tasks with a fixed output range, such as department routing, risk scoring, ranking candidates, or checking policy eligibility. Explaining a duplicate charge, outlining a refund, or asking for missing facts requires free-form language. Jev and a general LLM can serve different classification and explanation stages, but connecting them does not guarantee accuracy or speed.

Map how each classification affects the next step. A value that selects a queue differs from one that automatically approves a refund in acceptable error and review requirements. Even with the same model, automation scope depends on downstream permissions.

Some Qwen implementations read candidate-token logits directly; Jev-Omni uses a separately trained classification head over Gemma 4 12B IT representations. Both can avoid long text generation, but their scoring paths and model artifacts differ.

Define allowed choices, tie handling, and human review for uncertain cases before exposing an API³. Log the schema version because changing option labels can disconnect existing ground-truth labels.

TypeSafe keeps Jev's weights and training private. Local projects can read candidate logits from public Qwen checkpoints⁴ or prefill⁵ shared documents before branching into questions. They implement Jev-style decision outputs through separate approaches.

Models, prompts⁶, input processing, and parallelism can differ between implementations. Evaluate accuracy, probability calibration, and latency using the same data and conditions. Describe them as implementations aiming to reproduce Jev-style decisions rather than claiming they are equivalent to Jev.

To try this locally, start with the Qwen-based local Jev implementation comparison. It covers CPU⁷ experiments and projects for Mac and NVIDIA GPUs.

AutoTrust released autotrust/JEV-27B-VL to apply Jev-style decisions from text inquiries to images and screenshots. System 1 accepts images and text, then returns probabilities for yes/no, a choice among 2–256 options, or a 0–5 rating in one forward pass.

System 2 uses the Qwen3.8-27B backbone for ordinary image-input text and reasoning. The model is Apache-2.0 and contains 28B BF16⁸ weights. Official model card and 2026-09-30 announcement

Consider classifying a product photo as electronics, clothing, or home. Put the image and seller title in state, describe the classification question in question, and read the probability for each option. Use choice when you need a selected category, and noul when you need the probability that a property is true, such as whether an item is safe. Do not automate payment or permission decisions based only on a high probability; first evaluate errors and review thresholds on a human-labeled image set.

When the command completes, the JSON response includes the options and a probabilities array. For hf: command not found, check that the Hugging Face CLI⁹ is installed. If the image request is rejected, first check the state array’s {"image": URL} form and that the image URL is reachable. For a /v1/decide route error, check the server log to confirm serve_decide.py started in the required vLLM development environment. Official decision API and image input format

Include a gold action and an ‘unable to decide’ option for 100 tickets, then calculate exact-match accuracy, error types, and empty responses for each model. Repeat latency measurements by input length on the same device and record p50 and p95.

Route low-confidence, high-cost decisions to a person. Do not deploy based on speed alone; compare the cost of a wrong automated action with the review cost.

In a 2026-10-03 card update, AutoTrust reports completing 95% of 60 random multi-step tasks in headless Chromium, at about 0.26 seconds per click. Tasks involved three to seven clicks across shopping, settings, and email screens. The model received both numbered boxes and the text of screen elements. Under a separate condition with numbered boxes only, completion was 10%, so the 95% result is not a general GUI success rate from screenshots alone. Computer-use conditions in the official card

The robot-arm result is also a simulation, not a physical robot test: 75% completion across 20 random MuJoCo scenes, at about 240 ms per decision. Its control loop and environment differ from browser clicks, so the two latency figures are not directly comparable. The vLLM memory notes list 52 GB for BF16 weights plus about 17 GB for the KV cache¹⁰ at a 256K prompt, measured on a 183 GB B200. The card recommends starting at --max-model-len 131072 on an 80 GB GPU¹¹.