Is Jev open source?
Short answer: no. Jev is TypeSafe AI's proprietary, closed-weight System One model — you call it as a hosted API; the weights, training data, and exact architecture aren't released. What is open is the shape it popularized: a handful of open-weight, self-hostable decision models now answer the same typed question. Here's the honest breakdown of both.
There are two things people mean by 'is Jev open source': is the Jev model itself open (can I download the weights and run it?), and is there any open model that does what Jev does. The answers are different, so this page covers both.
Jev itself: closed-weight and hosted
Jev is proprietary. TypeSafe AI ships it as a hosted API — you send state plus questions and get back typed, calibrated answers — but it has not published the weights, the training data, or a full model card. Independent coverage of Jev's speed, cost, and accuracy figures traces them back to the company and its investors rather than an open checkpoint you can audit. So you can use Jev, but you can't self-host the real thing.
That's a deliberate trade. What makes Jev Jev isn't just the architecture (a non-autoregressive parallel sampler that emits a structured value in one forward pass) but the RLCD calibration — Reinforcement Learning for Calibrated Decisions — that makes its probabilities trustworthy. That calibration is the hard, closed part; an open reimplementation of the architecture doesn't inherit it.
What is open: Jev-like decision models you can self-host
Jev kicked off a category, and the open-weight side filled in fast. These are not Jev and not made by TypeSafe — they're separate open models that accept the same decision shape (a bounded answer space in, a typed result out). If your priority is open weights, data control, or zero per-call cost, these are the honest options:
- Laya — a 421M open-weight decision model (Convai, Apache-2.0) you can download and run locally, e.g. via Ollaya. Not RLCD-calibrated, so you validate confidence yourself.
- CLM-8B — an open candidate-ranking model from Stanford and NVIDIA; it scores candidates you supply rather than choosing against criteria you describe.
- Strands Decider 2B — a small open decision model released by AWS, aimed at cheap self-hosted gating.
- Clef — an open model that freezes a Qwen backbone and adds a decision head; Microsoft-Decision-1 similarly post-trains an existing open model (Qwen3.5-9B) for the job.
- Community reproductions of Jev's contract appeared within days of launch — architectural approximations built on ordinary LLMs, not weight-compatible with Jev. The community awesome-jev list is the cleanest map of these.
Closed Jev vs an open decision model
| Jev (TypeSafe) | Open decision models | |
|---|---|---|
| Weights | Closed — hosted API only | Open — download and self-host |
| Calibration | RLCD-calibrated probabilities | Not RLCD-calibrated — you validate it |
| Ops | None — you call an endpoint | You run and scale the model |
| Cost | Sub-cent per decision, output free | No per-call fee; you own the compute |
| Data control | Sent to the hosted API | Stays on your infra |
| Output shape | choice · score · noul + confidence | Same shape, varying quality |
If you still want Jev specifically
If it's Jev's calibrated hosted model you're after rather than open weights, you don't need the TypeSafe waitlist — our hosted endpoint serves the same choice / score / noul API with a jv_live_ key, billed per input token (output free) at well under a cent per decision. Try it free in the browser playground first, then grab a key. Exact numbers live on the pricing page.
FAQ
Is Jev open source?
No. Jev is TypeSafe AI's proprietary, closed-weight System One model — it's offered as a hosted API, and the weights, training data, and exact architecture aren't published. You can call Jev, but you can't download and self-host it.
Can I download the Jev weights or run it locally?
Not the real Jev — there's no public checkpoint. If you need something to run locally, use an open decision model such as Laya (self-hostable via Ollaya), Clef, or AWS's Strands Decider. They accept the same decision shape but aren't RLCD-calibrated, so you validate accuracy and confidence yourself.
Why isn't Jev open source?
The hard, valuable part of Jev is the RLCD calibration that makes its probabilities trustworthy, not just the parallel-sampling architecture. Open reimplementations can copy the contract and even the architecture, but they don't inherit that calibration — which is why TypeSafe keeps the weights closed.
Is there an open-source alternative to Jev?
Yes, several: Laya (421M, Apache-2.0), CLM-8B (Stanford/NVIDIA), AWS's Strands Decider, and Clef, plus community reproductions that rebuild Jev's contract on ordinary LLMs. None are made by TypeSafe. See the alternatives and awesome-jev pages for the full map.
What is OpenJev?
OpenJev is a community name used by more than one open reimplementation of Jev's decision contract — the best-known started under that name and was later renamed. They're architectural approximations built on existing LLMs, not weight-compatible with Jev and not made by TypeSafe. Treat them as separate community projects and check each repo's license yourself.
See also: Jev alternatives · awesome-jev list · CLM-8B vs Jev · Strands Decider vs Jev · Run a local model · Pricing · Get a key
Want the calibrated model, not the self-host project?
Call the same choice / score / noul API through our hosted endpoint — try it free in the browser, then grab a jv_live_ key. Sub-cent per decision, output free.