Decision model comparison · Julia 1 sources checked September 29, 2026
Jev vs Julia 1
Both turn a state, a question, and candidate answers into a structured decision. Jev is a hosted production API; Julia 1 is an open-weight model designed to run locally on a CPU.
Hosted API
Jev
TypeSafe · System One model
Use a managed decision model through an API without downloading weights, provisioning hardware, or operating an inference runtime.
Open weights
Julia 1
Supersonic Labs · 144.3M parameters
Download an Apache-2.0 checkpoint and run a focused decision model locally with Python, CPU inference, and caller-defined options.
Julia 1 is an independent project from Supersonic Labs. Jev AI is not affiliated with or endorsed by Supersonic Labs.
Published evaluation · September 24, 2026
Julia 1 is strong on focused choices, not every workload
Supersonic Labs reports a compact model that performs well on several classification pilots and stays close to the Jev reference on typed decisions. The spread across tasks matters more than a single headline score.
Typed decisions
2,000 questions
73.15% · 1,463 / 2,000
72.70% reference · +0.45 pp
AG News · 4 labels
100 pilot cases
94.00% · 94 / 100
91.00% reference · +3.00 pp
DAIR Emotion · 6 labels
100 pilot cases
86.00% · 86 / 100
48.00% reference · +38.00 pp
Banking77 pilot · 72 labels
100 pilot cases
64.00% · 64 / 100
87.00% reference · −23.00 pp
Julia 1's typed-decision breakdown is Choice 71.33% (428/600), Noul 80.67% (484/600), and Score 68.88% (551/800). Its reported MASSIVE scenario accuracy is 71.50% across 52 locales, including 86.75% for en-US.
These figures come from Julia 1's model card. The Jev values are supplied comparison references, not a new Jev run under the same environment. The pilots are signals, not guarantees for a new domain; test the exact questions and options in your own workflow.
Feature by feature
Same decision shape, different operating model
Both systems can express choice, score, and Boolean decisions. The meaningful differences begin with who hosts the model, how options are supplied, and how much infrastructure you want to own.
What it is
Hosted System One decision model
144.3M-parameter open-weight decision model built on mmBERT-small
Deployment
Call the production API; no model server to operate
Download the checkpoint and run the Python runtime on CPU or a compatible GPU
Question types
Noul, choice, and score, with multiple questions in one request
Choice, ordered score, and Boolean noul through one named-question interface
Options per choice
Use Jev's API limits and schema
Native calls accept 2–20 options; a router can narrow larger lists
Probabilities
Hosted probabilities for downstream decision thresholds
Full softmax probabilities in caller order; validate calibration for your task
Context and limits
Managed by the API contract
Runtime supports up to 8,192 combined tokens; historical benchmarks used 1,024
License
Proprietary hosted service
Apache-2.0 model artifacts and inference code; training pipeline is not released
Choosing a fit
Start with the constraint you need to solve
Choose Jev for the shortest path to production
Use it when you want typed decisions and probabilities in an API without managing weights, GPUs, model downloads, or runtime updates.
Choose Julia 1 for local control
Use it when open weights, CPU execution, offline deployment, or the ability to inspect and own the inference stack matters more than managed operations.
Evaluate both on your real decisions
They are decision models, not general chat systems. Clear options, representative state, and a held-out test set will tell you more than a broad benchmark headline.
Julia 1's model card says it is not a generative model and should not be expected to supply missing facts or perform long multi-step reasoning. Keep options distinct and validate any workflow that can trigger consequential actions.
Sources and scope
This page summarizes the Julia 1 model card and Supersonic Labs' launch notes, then places them next to Jev's hosted API model. Read the original materials for implementation details and the latest revisions.
Figures and repository details were checked on September 29, 2026. Results can change with model revisions, datasets, prompts, hardware, and serving configuration.