A practical comparison · JevBench v1.3.0 · Updated September 22, 2026
Jev vs OpenJev
Two decision systems, one compatible idea. Here is where hosted Jev and the open-source OpenJev server actually differ.
On this site
Jev 1.13.0
TypeSafe · System One model
74.4
JevBench composite · #1
The hosted decision model behind this playground.
Open source
OpenJev
razorback16 / Codiv · DiffusionGemma 26B-A4B
66.4
JevBench composite · #11
A Jev-compatible decision server you can host yourself or run through Codiv.
Where it runs
Hosted production API
Self-hosted on a 24GB GPU or Apple silicon
Request format
The original /v1/systemone
Wire-compatible with TypeSafe SDKs
Inputs
Text only
Text plus up to 8 images
Options per choice
Up to 255
Up to 128
Confidence
Calibrated probability
Entropy-derived confidence
Thinking mode
Not available
Available, with a quality / latency trade-off
OpenJev in this comparison means razorback16's DiffusionGemma-based project, not every project that has used the name.
Measured the same way
The benchmark tells a more useful story than a single score
JevBench v1.3.0 ran both systems through the same decision set. Jev leads on the balanced configuration; OpenJev becomes stronger when its thinking mode is enabled.
Composite score
A geometric mean of intelligence, calibration, speed, and cost.
Jev
74.4
OpenJev
66.4
Jev ranks #1; OpenJev ranks #11 in the published comparison.
Capability
Higher is better
Intelligence
How often it picks the right answer
85.7 / 79.2
Calibration
Whether 0.8 means about 80%
82.7 / 64.8
Speed
Measured response time
83.3 / 83.2
Accuracy by decision difficulty
Easy
72 straightforward cases
Standard
96 everyday cases
Judge
146 evaluation-style calls
Hard
220 genuinely ambiguous cases
In thinking mode, OpenJev reports 88.0 intelligence and 78.2% on the hard tier. That is a different operating point, not a like-for-like default configuration.
Figures are quoted from the comparison page and its published sources; validate them against your own workload before making a production decision.
Feature by feature
Same shape, different operating model
OpenJev deliberately speaks the same wire format as Jev. The meaningful differences begin after the request leaves your code.
| Attribute | Jev 1.13.0 | OpenJev |
|---|---|---|
| What it is | Hosted System One model from TypeSafe | Self-hosted decision server on DiffusionGemma 26B-A4B-it |
| License | Proprietary, hosted | Apache-2.0; weights from NVIDIA and Google |
| Request format | POST /v1/systemone | The same format; accepts openjev-latest and jev-latest |
| Question types | Noul, choice, and score — many in parallel | Noul, choice up to 128 options, and score over 2–10 levels |
| Inputs | Text, JSON objects, or arrays | Text plus up to 8 images per request |
| Off-schema answers | Typed answers only | Structurally impossible: the model reads answer slots |
| Uncertainty handling | Calibrated probability per answer | Re-reads up to four times when entropy is high, then averages |
| Hardware | None — it is an API call | 24GB NVIDIA GPU under vLLM, or Apple silicon with 16GB free under MLX |
| Cost to run | Metered in Jev AI credits | About $0.066 per 1,000 decisions on a rented GPU, or free on Codiv |
The trade-offs
What each one is better at
There is no universal winner. The right choice depends on whether you value managed access, calibrated confidence, images, or control over the runtime.
Choose Jev for
- A confidence value you can use without fitting a calibration layer yourself.
- Stronger hard-tier accuracy in the ranked configuration: 74.1% versus 65.5%.
- Up to 255 options per choice, with no GPU, vLLM, or quantized checkpoint to operate.
- One stable hosted configuration that is easy to pin and put behind production logic.
- A fast path from an experiment to a real API request.
Choose OpenJev for
- A genuine drop-in for TypeSafe's SDKs: change the base URL and keep the request shape.
- Image input, with up to eight images in one decision request.
- A thinking mode that can outperform the default configuration on difficult cases.
- A structurally typed answer path where the model does not write into the answer slots.
- Apache-2.0 code and weights that can run inside your own network or on Apple silicon.
From the repository
What razorback16's OpenJev actually ships
OpenJev is more than a model checkpoint: it is a small decision server around DiffusionGemma with vLLM and MLX backends.
Open the GitHub repositoryCompatible API
POST /v1/systemone, plus OpenAI-style /v1/chat/completions
Backends
vLLM for NVIDIA GPUs; MLX for Apple silicon
Deployment
Docker or local Python process; free hosted access is available through Codiv
License
Apache-2.0
OpenJev is an independent project and is not affiliated with or endorsed by TypeSafe AI.
So which should you use?
Start with the constraint that matters most
Use the decision model that matches your deployment boundary and the kind of uncertainty your product needs to handle.
Pick Jev if
- You branch on probability and want it to mean something out of the box.
- You do not want to operate a 26B diffusion model or tune denoising steps.
- Your inputs are text and you want one stable, managed configuration.
Pick OpenJev if
- Your decision involves an image or benefits from a reasoning budget.
- You already run GPUs, or the data must stay inside your network.
- You will validate confidence thresholds on your own labelled data.
The fastest answer is your own test set
Try Jev on the cases you are actually arguing about
Run a few awkward, borderline decisions in the playground, then compare them with the OpenJev setup that fits your hardware.
FAQ
Jev vs OpenJev questions
Can I point the TypeSafe SDK at OpenJev?+
Yes, for razorback16's server. It implements the /v1/systemone wire format and accepts Jev-compatible model aliases, so changing the base URL is the main integration step.
Is OpenJev as accurate as Jev?+
In the ranked configuration, OpenJev trails Jev on the hard tier and calibration. Its thinking mode is a different configuration and reports higher intelligence on the published hard-case results.
What hardware does OpenJev need?+
The repository documents an NVIDIA setup with about 24GB of GPU memory, or an Apple silicon setup with about 16GB free for the 4-bit MLX weights. Codiv also provides hosted access.
Does OpenJev replace Jev?+
It can replace the endpoint for some workloads, especially when images, self-hosting, or Apache-2.0 matter. It is not a drop-in replacement for the operational guarantees of a managed hosted API.
OpenJev and the other projects named here belong to their respective owners. Figures are quoted from public sources as checked on September 22, 2026.