shipwithjev

Blog / Questions / FIG. 53

Can Jev Run Locally? No, and Here's the Decision Tree That Follows

Jev is hosted-only: no local weights, no on-device inference. What that means for offline, edge, and privacy cases, and the honest alternatives for each.

No. Jev is a hosted model accessed through TypeSafe's platform (via the Vercel AI Gateway, per ecosystem documentation), the weights aren't published, and there is no local, on-device, or air-gapped inference path. Every verdict is an API call. That's the fact; here's the decision tree it creates, because "I wanted local" means four different things.

If you meant privacy: the hosted reality is softer than it sounds, because minimization is architectural: send fields not records, extract locally and ship only text per the screenshot-free pattern, and the data that never leaves can't leak. Most "must run local" privacy cases are actually "must minimize" cases wearing a stricter costume.

If you meant offline or edge: design for the seam. Robotics and device builds already split this way, local reflexes on firmware, remote verdicts for the deliberate layer, and every pipeline needs its fallback policy (queue, degrade to rules, route to human) for the disconnected minutes regardless.

If you meant cost at extreme volume, or hard residency/air-gap requirements: that's the one branch where hosted genuinely can't follow, and the honest answer is the open pattern: TypeLLM over models you serve yourself, trading reported per-verdict prices for sovereignty and a GPU bill. Your question sets port, which is what makes the branch survivable.

If you meant latency: check the receipts before assuming; game loops sustain ~10 verdicts a second over the wire, which covers more real-time cases than intuition predicts. Single-digit-millisecond floors stay with trained local models, as they always did.

Frequently asked questions

Is there any on-device Jev, even reduced?

None documented; ecosystem experiments pair local perception with remote verdicts instead. If that changes, the robotics page updates first.

How do I handle API downtime for a critical pipeline?

A pre-decided fallback per pipeline (queue-and-wait, rules floor, human lane), per the production guide: the policy chosen calmly beats the one improvised at 3 a.m.

Does hosted-only make Jev unsuitable for regulated data?

It makes the vendor review mandatory (TypeSafe's residency and retention commitments) and minimization non-optional; many regulated pipelines pass on that basis, and the ones that can't have the TypeLLM branch.

Numbers throughout are as reported by the build authors, not verified by shipwithjev. Code-shaped examples are pseudocode; the official docs live at docs.typesafe.ai.