/ 01
Short answer
You cannot currently run TypeSafe’s official Jev model on your own hardware. TypeSafe exposes Jev as a hosted model; its public SDKs send requests to that service and are not local inference runtimes. You can run your application, SDK, gateway or workflow locally while the decision still runs remotely. Or you can run a separate open model locally and implement a Jev-like typed decision pattern. Those are useful options, but neither makes the official Jev weights self-hosted. See the TypeSafe model documentation and the current Jev API guide.
/ 02
What “self-hosted” means for Jev
| What runs on your infrastructure | Can you do it? | What it means |
|---|---|---|
| Official Jev model weights and inference | No public release identified | The actual Jev decision still runs on TypeSafe infrastructure. |
| Your app, SDK, gateway or workflow | Yes | You control this code and where it runs, but it calls a hosted inference provider. A local proxy does not move the model. |
| A separate local model with typed decisions | Yes | Run downloadable weights and serving software yourself, then map scores into Choice-, Score- or Noul-shaped results. It is a different evaluator, not Jev. |
“Local,” “private,” “edge,” and “self-hosted” describe different controls. If your requirement is that prompts never leave your network, a locally hosted client or gateway alone does not satisfy it; inference must also run on a model you control.
/ 03
The three claims you’ll see online
| Source | What the page says | How to read it |
|---|---|---|
| Failproof AI: Can you self-host Jev? | No official downloadable weights or self-hosted inference package was identified; SDKs and provider routes are hosted access. | Consistent with the distinction above. “Not identified” is a dated public-source check, not a promise about future releases. |
| AI Profit Boardroom: Is Jev Open Source? | The real Jev model cannot run locally because its weights are not published; local glue and a local approximation can. | The headline answer is “no” for official Jev. The article also distinguishes local orchestration from local inference. |
| Daily Dose of Data Science: Build your own Jev (100% local) | Uses SGLang scoring with open models to reproduce fixed-choice scoring locally. | The title is shorthand: the article says it recreates the inference path, not the complete Jev system, and specifically excludes Jev’s training and calibration. |
These claims are less contradictory than their titles suggest. Two discuss whether the official model can be downloaded; the third demonstrates a local approximation of one decision mechanism. None establishes that TypeSafe’s Jev weights can be self-hosted.
/ 04
What you can run locally
A local substitute can use a model with downloadable weights, a local inference server, and application code that supplies a state plus a finite set of permitted outcomes. For Choice-like decisions, one implementation scores candidate labels, normalizes the scores into a distribution, and maps the winning label back to your application. Code can define ordered levels for Score-like output and a binary decision for Noul-like output.
This is an implementation pattern, not a drop-in equivalence claim. The base model, scoring method, probability calibration, supported context, latency, and failure modes differ. Measure it on your own labeled examples, set thresholds for that evaluator, and keep a human-review path for uncertain cases. The Daily Dose article links to its worked SGLang scoring approach; inspect project and model licenses before adopting a particular stack.
/ 05
What only the hosted API gives you
The hosted route gives you access to TypeSafe’s Jev model and its trained decision behavior without operating model weights or an inference fleet. TypeSafe’s published direct-API figures are up to 250,000 tokens per second and 1,200 requests per minute; the currently listed price is $0.042 per million input tokens. The provider says rate limits can change, and a gateway may apply separate limits or billing. These are hosted-service figures, not capacity figures for local models. Check the current TypeSafe model page and our Jev limitations page for the dated context, route distinction, 64k/32k context descriptions, and pricing notes.
For endpoint setup see the API guide, and for VerdictKit plan costs see pricing. Self-hosting trades usage billing for hardware, capacity planning, updates, monitoring, and on-call responsibility.
/ 06
Decision table
| Your requirement | Starting point | Trade-off |
|---|---|---|
| Use the official Jev model and its hosted behavior | Call the hosted API or a supported provider route | Your request goes to the selected inference provider; review its data, retention, cost and rate-limit terms. |
| Keep app code and credentials under your control | Run your app or gateway in your environment and call a hosted model | Controls the client and network path; it does not make model inference local. |
| Keep prompts on your hardware or work offline | Run a separately licensed local model and typed decision layer | Local data path and independent capacity; you own serving and must re-evaluate quality and calibration. |
| Reproduce the typed-decision interface for a prototype | Try an open local scoring implementation | Fast way to test the interface; output semantics and quality are not automatically equivalent to Jev. |
| Choose based on workload economics | Benchmark representative traffic against hosted and local options | Include accelerator utilization, peak load, engineering time, monitoring and fallback costs. |
/ 07
Community projects
For community context, see the LocalJev? Reddit discussion and the independent openjev-sglang GitHub project. Verify each project’s source, license, model artifact, and outbound network behavior before using it with sensitive data.
/ FAQ
Frequently asked questions
Can I self-host the official Jev model?
No public downloadable weights or self-hosted Jev inference package are identified in the public materials reviewed for this page. Recheck TypeSafe’s documentation for changes.
Can I use Jev locally through an SDK, Cloudflare, or a gateway?
You can run the SDK, app, or your own gateway locally, but if it calls a hosted Jev endpoint, the inference is still remote. A managed Cloudflare or other provider route is also hosted inference.
Is the “100% local Jev” tutorial actually Jev?
It recreates a Jev-like fixed-answer scoring path with open models running locally. The tutorial says it does not reproduce the complete Jev system, including its training and calibration.
Is there a self-hosted Jev discussion on Reddit?
Yes. A LocalJev? Reddit discussion covers running the decision pattern with local models. A Reddit post is not an official weights release; check any linked repository and license directly.
Is there a Jev self-hosting project on GitHub?
There are independent Jev-inspired and compatible implementations. One local server project uses an open model; it is separate from TypeSafe’s Jev weights. Review the openjev-sglang repository linked in the community pointers above for its current code, requirements, model license and limitations before adopting it.
Should I replace hosted Jev with a local model in production?
Only after a side-by-side evaluation on representative labeled data. Recheck calibration, quality, latency under peak concurrency, security, model license and operating costs; do not carry Jev thresholds over without validating them.
Test the hosted model first — the playground runs Choice, Score and Noul in your browser.