OpenJev is a brand-new open-source approximation of TypeSafe's Jev — the SiliconLabAI/OpenJev GitHub project rebuilds Jev's decision contract (fixed answer spaces, parallel scoring, probabilities with confidence) on top of ordinary chat models, so you can run the System One pattern with any API key, no waitlist.
📺 Watch: Jev AI: 10 INSANE Use Cases
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
That last part is the headline. Real Jev sits behind an early-access queue at console.typesafe.ai, so most people reading about it cannot actually touch it yet. OpenJev flips that overnight: point it at whatever model you already pay for — or one you host yourself — and the pattern of asking a question and getting back a decision with a confidence score, instead of a rambling paragraph, is yours to experiment with today.
Everything in this article comes straight from the OpenJev README on GitHub. The repo is live, the first commit landed on 20 September 2026, and it had 32 stars at the time of writing. I have not tested it myself — it is days old — so treat this as a guided tour of what the project claims, not a hands-on review. And if you need the backstory on what Jev actually is, start with my full Jev AI guide first.
What OpenJev Actually Is
The README pitches OpenJev as an open approximation of TypeSafe Jev — a System One style decision engine — and the GitHub About line simply calls it OpenSource Jev. Both phrasings tell you exactly what this is and is not: an independent rebuild of the idea, not leaked weights and not an official TypeSafe release. Real Jev remains closed, a question I unpacked properly in is Jev open source.
The core idea is the interesting part. The normal way to get structured decisions out of an LLM is one big call asking the model to return a giant JSON blob, and anyone who has shipped that knows how flaky it gets. OpenJev replaces it with a parallel sampler. You declare a fixed answer space up front, so there is no free-form text generation at all. Each option gets scored independently against the same state. The scores get normalised from logits into a softmax probability distribution. And every question you send runs in parallel.
According to the README, that mirrors how Jev itself is described publicly: parallel evaluation over a declared answer set rather than sequential token generation. OpenJev copies the contract, not the model.
OpenJev GitHub
The repo lives at SiliconLabAI/OpenJev on GitHub, and freshness is the story: first commit on 20 September 2026, 32 stars at the time of writing. That cuts both ways. You get to explore before the crowd shows up, and you also inherit every rough edge of week-one software. Worth noting too: the README shows no licence at the time of writing, which matters the moment you want to build anything commercial on top of it.
The quick start is three steps. Run npm install, then npm run dev, then open http://localhost:3001 in your browser.
Configuration stays small. Four fields: mode (parallel by default, with a oneshot fallback), model (the README's examples are gpt-4o-mini and qwen-3.8-27b), base url (any provider you like, or leave it empty for OpenAI), and api key. Or skip the key field entirely and export OPENAI_API_KEY before you run it.
Drops like this are exactly what I flag inside AI Profit Boardroom the day they land — 3,000+ members get the new-repo alerts, the working configs, and the honest follow-up when something turns out half-baked, long before any of it reaches the blog.
How the Parallel Sampler Works
The scoring trick is beautifully simple, and the README lays it out in plain terms. For every candidate answer, the model gets one tiny prompt that boils down to asking how likely it is that this specific option is the correct one. It returns a probability. Those independent probabilities get treated as logits, and a softmax turns them into a proper distribution that sums to one.
Each question type then reads that distribution its own way. A choice takes the argmax — the highest-probability option wins — with confidence coming from how far the top answer leads the rest. A score takes the expected value across its ordered levels. A noul is a single yes/no probability, no distribution needed.
Why do all this instead of one big structured call? The README's own comparison makes the case:
| Mode | How it works | Failure behaviour |
|---|---|---|
| Parallel (default) | One tiny probability call per option, then a softmax across the results | Failures are rare — each call is tiny and tightly constrained |
| Oneshot | One big structured JSON response covering everything at once | Models drop keys, invent labels, or return invalid JSON |
In concrete terms, a choice with four options fires four parallel micro-calls, while a yes/no noul question needs exactly one. You trade a handful of extra requests for answers that almost never break — and because the sampler runs everything in parallel, you barely pay for it in wall-clock time.
The API Shape, In Plain English
You do not need a code sample to understand the API. You POST a state — the situation you want judged — plus a list of questions to an evaluate endpoint. The question types mirror Jev exactly: choice, where every option carries its own criteria description; score, which works across ordered levels; and noul, a straight yes/no probability.
Back comes the pick, a confidence figure, the full probability distribution across every option, token usage, and metadata covering latency and how many parallel calls actually ran. The README's example shows nine parallel calls completing in 620ms on gpt-4o-mini — a respectable number for stacked micro-calls on an ordinary chat model.
If that shape feels familiar, it should. It is the same contract I walked through in my Jev AI API guide, and that is the strategic point: learn it once against OpenJev, then swap in the real endpoint later without rewriting your application logic.
Not Real Jev — The Honest Differences
Credit to the maintainers: the README does not oversell. Its own words:
"Not real Jev."
Here is what that means in practice. OpenJev uses ordinary chat LLMs as micro-scorers. Real Jev is a specialised System One model — RLCD training, a custom parallel sampler built into the model itself, and responses in roughly 70-500ms. OpenJev is an open architectural approximation of the contract, not a weight-compatible reimplementation, so expect ordinary-LLM latency and ordinary-LLM cost on every micro-call, multiplied by however many options your questions carry. For the full picture of what makes the real model different under the hood, read my Jev AI model breakdown.
It is the same lesson Goldie Bench keeps teaching: the interface and the weights are separate achievements. You can clone a contract in a weekend. The speed and cost profile is the part that took a research team.
Who Should Use Which
Based purely on what the README claims, the split looks like this:
- Use OpenJev to prototype the pattern. Any model, any provider, right now. The base url field means the qwen-3.8-27b example can run against your own self-hosted server, which makes the Jev pattern — decisions with confidence attached — literally free to experiment with today.
- Use real Jev for production speed and cost. At the established pricing of $0.042 per million input tokens with output free, plus those 70-500ms responses, the real model stays the frontier. Early access runs through console.typesafe.ai, and I covered the access routes in how to use Jev AI for free.
The smart play is both. Build against the contract with OpenJev now, swap the endpoint when your Jev access lands. And if you want help turning this pattern into an actual revenue workflow instead of a weekend demo, book a free strategy session and we will map it onto your stack together.
Read the Code Before You Install Anything
My standard caveat, and it counts double here. I have not tested OpenJev. Nobody has meaningfully audited it. The first commit is days old, the star count is 32, and the README shows no licence. None of that makes it malicious — it makes it a brand-new GitHub repo, which is precisely the category of software you read through before you run npm install on a machine you care about. A parallel sampler this focused should be a small codebase, so skimming it is a twenty-minute job. It is the same discipline I push in the Agent OS guide: tools earn trust through inspection, not star counts.
OpenJev FAQ
Is OpenJev the same as Jev?
No. OpenJev approximates Jev's decision contract — fixed answer spaces, parallel scoring, softmax probabilities — using ordinary chat models as the scorers. Real Jev is a specialised System One model with RLCD training and a custom sampler answering in roughly 70-500ms. Same interface idea, completely different engine underneath.
Does OpenJev work with local models?
Yes, according to the README. The base url field accepts any provider endpoint, and qwen-3.8-27b appears alongside gpt-4o-mini in the example config. Point OpenJev at a self-hosted server and the parallel sampler runs entirely on hardware you control, with no per-token bill attached.
Is OpenJev free?
The code costs nothing to clone and run, though remember the README shows no licence at the time of writing. Your real cost is inference: every option in a choice question is its own micro-call, so you pay your provider's normal token rates — or nothing beyond electricity on a local model. Real Jev, for comparison, charges $0.042 per million input tokens with output free.
Get Ahead of This One
A repo this new — 32 stars, first commit on 20 September 2026 — is exactly the window where paying attention pays off. Whether OpenJev becomes the standard way people prototype System One decision engines or gets superseded next month, the pattern it teaches — declared answer spaces, parallel probability scoring, confidence you can act on — is worth learning right now.
Inside AI Profit Boardroom — $69/mo locked in, normally $110 — 3,000+ members turn drops like OpenJev into working systems every week, with the configs and the failures shared in the open. Prefer a one-on-one look at your setup instead? Grab a free strategy session, bring your use case, and we will figure out whether the Jev pattern belongs in it.
