agentos.guide › blog

Can You Run Jev Locally? The Honest Answer, Plus The Local Setup I Use

By Julian Goldie · 5 October 2026 · agentos.guide

No, you cannot run Jev locally, because TypeSafe has not released the model weights.

You can run a Jev-shaped decision model locally, though, and I have one running on my own Mac right now.

The question I keep getting is: can I run Jev locally so my data never leaves my machine?

Here is what is actually possible today, the real numbers from my own tests, and how I would wire a local decision model into an Agent OS without losing accuracy.

Why you cannot run Jev locally today

TypeSafe has not released Jev's weights, so there is nothing to download.

Jev only exists as a hosted service that you call through TypeSafe's API, OpenCode Zen, Vercel AI Gateway or OpenRouter.

The SDKs are open source under the MIT licence, but an SDK is just the code that talks to the hosted model.

So any page promising a Jev download is either selling you something else or confusing the SDK with the model.

I want to be really clear about that, because it saves you an afternoon of hunting.

🔥 Want the exact local decision model setup I run inside my Agent OS?

Inside the AI Profit Boardroom you get the installable Agent OS, step-by-step video tutorials, four coaching calls a week and 3,400+ members building real automations.

→ Get access here

Why people want to run Jev locally in the first place

The reasons I hear from members always come down to three worries.

The first worry is privacy, because a decision layer sees every email, ticket and lead that flows through your business.

The second worry is lock-in, because a hosted model can change its price, its terms or its free tier overnight.

The third worry is speed and cost at volume, because an agent can make thousands of tiny decisions a day.

Those are fair worries, and the good news is that a local, Jev-shaped decision model now exists.

It just is not Jev.

What "running Jev locally" really means

When people ask this, they usually mean one of two different things.

Some people want the exact Jev model on their own hardware, and that is not possible today.

Other people want the Jev behaviour on their own hardware, which means typed questions in and probabilities out with no data leaving the building.

The second one is very possible now.

A Jev-style model answers yes-or-no, pick-one and score questions with a confidence number, exactly like Jev does.

If it accepts the same request shape, your agents cannot tell the difference except in the answers.

That is the version of "local Jev" I will show you.

The local Jev alternatives, side by side

The important line in that table is "Jev-API compatible".

It means code you write for Jev can point at the local model by changing the URL, the key and the model name.

That is the property that makes a local swap realistic inside an Agent OS.

How I run a Jev-style model locally on my Mac

Laya is the one I have actually installed and tested on my own machine, so I will walk you through that route.

• Install it with pip install laya inside a fresh Python virtual environment.
• Expect the first run to download the model checkpoints, which took a couple of minutes on my machine.
• Ask it the same kinds of questions you would ask Jev, such as a department choice, an urgency score or a yes-or-no refund check.
• Use Router(preload=True) if you will see more than one language, because the lazy router reloaded a model on every language switch.

On my Mac, one English question took about 20 ms, and a batch of ten questions came to about 3.9 ms each.

With every model preloaded, mixed-language messages came back in about 19 ms each.

Without preloading, each language switch took around 20 seconds, which is the gotcha that catches people out.

The honest catch with running a Jev alternative locally

Fast is not the same as right, and this is where I want you to slow down.

On my 60-email test, hosted Jev got 54 right.

Laya with no training got 28 with its English model and 36 with its typed-decisions model.

On 24 made-up contact-form leads, Laya got 16 and Jev got 24.

On picking which of five websites an article brief belongs to, Laya got 8 of 10 and Jev got 10 of 10.

The good news is that Laya did not pretend to be sure when it was guessing.

None of its 60 email answers came back at 0.85 confidence or higher, so the confidence number was honest even when the answer was not.

The Laya team's own published numbers show it beating Jev after training on their benchmark, but I have not run that training myself.

So treat an untrained local model as a fast base that needs specialising, not as a plug-and-play Jev replacement.

Where Clef fits if you have the hardware

Cloudflare released Clef and Clef-flash on 1 October 2026 with open weights under Apache 2.0.

Cloudflare says both are fully Jev-API compatible and that Clef has a 64k context window, against Jev's 32k.

On Cloudflare's own published benchmarks, Clef scored 94.20 on BANKING77 against Jev's 79.74.

Cloudflare also reports Clef-flash at a 38.8 ms median latency, against 524.1 ms for Jev in its tests.

Those are Cloudflare's numbers on Cloudflare's chosen tests, so check them on your own data.

I have not run Clef on my own Mac yet, so I am not going to give you a speed number I have not measured.

The 27B model is a serious download for a laptop, so the 9B Clef-flash is the sensible first test for most people.

How to wire a local Jev alternative into your Agent OS

This is where the Agent OS approach really pays off.

In my setup, the decision layer is a slot, not a brand.

The agents ask the slot a typed question, and the slot decides which model answers it.

• Put short, private, high-volume questions with fewer than 20 options on the local model.
• Send long option lists, such as 77 categories, to hosted Jev, because on the 77-category test covered in my Laya guide, Jev scored 0.870 and Laya scored 0.425.
• Keep one config value for the decision endpoint so you can swap Laya, Clef or Jev without touching your agents.
• Log the confidence number on every call, so you can see which lane is guessing.
• Send anything under your confidence line to a human or to a bigger model for a second opinion.

That gives you privacy on the bulk of your traffic and accuracy where it matters.

It also means a price change at any single provider becomes a config edit instead of a rebuild.

A simple routing rule you can copy

If the question has fewer than 20 options and the data is private, ask the local model first.

If the local confidence comes back low, ask hosted Jev the same question.

If both are unsure, a human gets the job.

That three-step rule is boring, and boring is exactly what you want in a decision layer.

Step by step: a local decision layer in one afternoon

Here is the order I would follow if I were starting from scratch today.

• Write down the five small decisions your agents make most often, such as which folder, which department, how urgent and whether a task is finished.
• Turn each one into a typed question with a short instruction and, for choice questions, an "other" option.
• Collect 50 to 100 real examples for each question and note the right answer yourself.
• Run those examples through hosted Jev first, so you have a baseline score to beat.
• Run the same examples through your local model and compare the scores and the confidence numbers.
• Keep the local model for every question where it scores close to Jev, and leave the rest on the hosted lane.

That last step is the whole game.

You are not trying to replace Jev everywhere on day one.

You are moving the easy, private, high-volume questions onto your own machine and leaving the hard ones where the accuracy is.

Common mistakes when people try to run Jev locally

The first mistake is downloading something that claims to be Jev, because the real weights have never been released.

The second mistake is trusting an untrained local model's top answer without checking it on your own examples.

The third mistake is forgetting to preload the models, which turns a 20-millisecond answer into a 20-second wait every time the language changes.

The fourth mistake is putting a 27B model on a laptop that cannot hold it, when a smaller model would answer your questions just as well.

The fifth mistake is hard-coding the model into every agent, which makes the next swap painful.

Should you bother running Jev locally at all?

If you are a solo operator sorting a few hundred emails a week, hosted Jev is so cheap that local is mostly about privacy.

TypeSafe lists it at $0.042 per million input tokens with free output, and my calls through OpenRouter came to about $0.00002 each.

If you handle client data for an agency, the privacy case for a local model gets much stronger.

If you run agents all day across lots of sites, the lock-in case gets stronger too.

My view is simple: learn the decision-layer pattern once, and the model underneath becomes a swappable part.

If you want the full decision-layer setup inside the Agent OS, with tutorials as each new model lands, it is all inside the AI Profit Boardroom.

FAQ

Can I run Jev locally?

No. TypeSafe has not published Jev's weights, so it only runs as a hosted service. You can run open, Jev-style models such as Laya, Clef or Clef-flash on your own hardware instead.

Is Jev open source?

The model is not. TypeSafe's JavaScript and Python SDKs are MIT-licensed, but they only talk to the hosted model.

What is the best local Jev alternative?

Laya is the one I have installed and tested on my Mac, and it is very fast but needs training to match Jev's accuracy. Cloudflare's Clef and Clef-flash are newer open-weight options that Cloudflare says are fully Jev-API compatible.

How fast is a local Jev alternative?

On my Mac, Laya answered one English question in about 20 ms, and preloaded mixed-language questions in about 19 ms. Without preloading, a language switch took around 20 seconds.

Will a local model be as accurate as Jev?

Not out of the box in my tests. Untrained Laya scored 28 and 36 out of 60 on my email test, where hosted Jev scored 54, so plan to train it or keep Jev as a fallback for hard questions.

The bottom line

The short answer is no, but a local, Jev-compatible decision layer is now a real option if you test it on your own data first.

That is the honest state of play if you want to run Jev locally in October 2026.

Related guides on agentos.guide

→ Laya AI: The FREE Jev AI Alternative (7X Faster)

→ Jev AI Model Routing Changes Everything

→ 3 Free Ways To Use Jev AI Right Now

📺 Video notes + links to the tools 👉

🎥 Learn how I make these videos 👉

🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉

← all guides on agentos.guide