Can I Run Jev Locally? The Honest Agency Answer (Free)

Can I run Jev locally? No, you can't run Jev locally, because TypeSafe only offers it as a hosted service with no downloadable weights, but you can run a free Jev-compatible model called Laya on your own machine and keep client data in-house.

That's the short version for anyone running a business.

The longer version matters more, because "local" and "free" sound great until the wrong lead lands in the junk pile.

I've tested both sides of this on my own Mac, with real Jev calls and real Laya runs.

So in this post I'll talk about it the way an agency owner should, in terms of cost, client data, accuracy and where each option earns its keep.

Why Business Owners Keep Asking If They Can Run Jev Locally

I get why this question keeps coming up.

Jev is a decision model from TypeSafe AI that launched on 15 September 2026.

It doesn't write emails or blog posts.

It reads something like a lead, a ticket or an email, picks one answer from a list you give it, and tells you how sure it is.

That's exactly the boring sorting work that eats a small team's day.

So the moment people see it, they ask three business questions.

The first is whether they can stop paying per question.

The second is whether they can keep client emails and lead data off somebody else's servers.

The third is whether a local version is good enough to trust with money on the line.

Let me answer each of those honestly.

The Straight Answer: Jev Is Hosted Only

TypeSafe's own site talks about early access, a console and a price of $42 per billion input tokens.

It doesn't mention downloadable weights, an on-premise version or an offline build.

Every way I've used Jev, whether through OpenRouter, Vercel AI Gateway or OpenCode Zen, sends the request to a hosted model.

So if a client asks you to "install Jev on our server", the honest reply is that nobody can do that today.

What you can install is a model that does the same job and accepts the same request format.

That's where Laya comes in.

What Laya Is, In Business Terms

Laya is a free, open-source decision model from a small team called Convai Innovations.

It's licensed under Apache 2.0, which means you can use it commercially, change it and run it on your own machines.

It ships as three small models behind one Router.

One handles English, one handles over 100 languages, and one was trained on business paperwork like invoices and tickets.

The Router looks at each message first and sends it to the right model automatically.

Its local server copies Jev's request format at POST /v1/systemone, according to the Laya README.

In plain terms, that means software built for Jev can be pointed at Laya by swapping one web address.

The Cost Picture: Local vs Hosted Jev

Let's talk money, because that's why most of you are reading this.

Option	What you pay	Where client data goes	Notes
Jev via OpenRouter or Vercel	$0.042 per million input tokens, output free	To the hosted provider	Vercel's free promo ended on 25 September 2026
Jev via OpenCode Zen (`jev-1.13-free`)	Free for a limited time	To OpenCode's hosted service	Read the Zen data policy before sending anything sensitive
jevplayground.com	Free, no signup	To an independent test bench	Good for testing, not for client work
Laya on your own machine	Free per question	Stays on your machine	You pay in setup time and, ideally, training time

Here's the thing most people miss.

Hosted Jev is already very cheap per question.

Builders reported sorting 500 emails for about $0.035 during launch week, although that's a self-report rather than an independent test.

So for most small businesses, the reason to go local isn't the bill.

The real reasons are client data, speed and not being tied to a free window that can close.

The Client Data Argument Is The Strongest One

If you run an agency, you're handling other people's inboxes, leads and customer complaints.

Every hosted call sends that text to a third party.

With Laya running on your own machine, the text never leaves the building.

That's a much easier conversation with a cautious client than explaining someone else's data policy.

My guide on the free Jev doors flagged this too.

Free windows sometimes come with different terms, so it's worth two minutes reading the data policy before you push anything sensitive through them.

🔥 Want the exact decision-layer setup I'm building for my own agency? Inside the AI Profit Boardroom, I share the lead-sorting, ticket-triage and content-routing walkthroughs as I build them, with Laya or Jev sitting in front of the main agent. Plus weekly coaching calls + 3,400+ members building real automations. → Get access here

What I Actually Saw When I Tested Laya On Business Tasks

I ran Laya 0.3.4 on my own Mac on 21 September, side by side with real Jev.

All the test data was made up, so no real customers were involved.

Here's what happened on three jobs every agency recognises.

Sorting Contact-Form Leads

I fed both tools 24 contact-form messages and asked whether each was a lead, a vendor or noise.

Jev got all 24 right at about 392 ms each.

Untrained Laya got 16 right at about 57 ms each.

That's fast, but a third of your leads in the wrong pile is not acceptable for a real business.

Sorting An Inbox

I ran the same 60 test emails Jev had sorted for me two days earlier.

Jev got 54 of 60 right.

Untrained Laya got 36 with one model and 28 with the other.

The good news was that Laya never once claimed to be 85% sure on that test, so it wasn't confidently wrong.

Routing Content Briefs To The Right Site

I gave both tools 10 article briefs and five possible websites to publish on.

Jev picked the right site all 10 times.

Untrained Laya picked the right site 8 times, in milliseconds.

That's close enough to be useful as a first pass with a human check.

The Training Step Is Where The Money Is

Those untrained scores are not the end of the story.

The version of Laya trained on business paperwork beat Jev on a 2,000-choice business test, 0.766 against 0.727.

To be fair, that model was trained on the practice version of that exact test, which the Laya team was open about.

But it shows what happens when you train it on the kind of text you actually see.

The Laya team ships a training notebook that runs on Kaggle's free GPUs.

It builds your training data, trains the model, tunes the sureness numbers and saves the result.

My estimate from the guide was four to five hours, with no card needed.

For an agency, that's the play.

You spend one afternoon training it on your own leads or tickets, and then every decision after that is free and stays on your machine.

When Paying For Hosted Jev Is The Smarter Business Move

I'm not going to pretend local always wins.

Here's when I'd keep paying for Jev.

You have a big list of options. On a 77-category banking test, Jev scored 0.870 and Laya scored 0.425, because Laya squeezes long option lists together.
You don't have time to train. Untrained Laya was clearly less accurate on my lead and inbox tests.
The decision touches money. If a wrong answer costs you a sale, accuracy matters more than milliseconds.
You're still testing the idea. A free hosted door gets you running in seconds without any install.

My rule of thumb is simple.

Under 20 answer options, test Laya.

Over 50, stay on Jev or split the question into two smaller steps.

When Running Local Jev Pays Off

Here's when I'd go local with Laya.

Your clients are strict about data. Nothing leaves your machine, which is a strong selling point.
You have customers in several languages. The Router sends each message to the right model, and the Laya team reported 45 of 51 languages working well with it against only 23 without it.
You make thousands of tiny decisions a day. Laya answered in about 20 ms per question on my Mac, against Jev's published 236 to 276 ms.
You've trained it on your own examples. That's when it stops being a demo and starts being an asset.

Ways To Use A Local Decision Layer In Your Business

The real value isn't the model itself.

It's putting a fast, cheap decision step in front of everything else.

Here are the jobs I'd point it at first.

Lead triage decides whether a form fill is a hot lead, a vendor pitch or spam, and sends only the hot ones to your sales person.
Ticket routing sends each support email to billing, technical or sales, with a sureness number on every call.
Churn flags ask a yes-or-no question about whether a customer is threatening to leave, so a human gets it first.
Content routing decides which of your websites should publish a new brief.
Agent routing scores how hard a task is so the cheap model handles the easy work and the expensive one only gets the hard jobs.

If you're new to building these systems for clients, my guide on how to start an AI agency covers the business side.

For more ideas on where to plug automation into a small company, see my AI automation for small business breakdown.

The One Risk That Can Quietly Cost You Money

There's a trap with every one of these tools, local or hosted.

The sureness number only means something on text the model has seen before.

The Laya team's own write-up showed an English model scoring zero right on Khmer text while saying it was 95% sure.

On the emotion test, Jev gave the right answer a 0% chance 16% of the time.

If your rule is "send angry customers to a human when there's over a 20% chance", that means roughly one in six angry customers slips straight past.

So before you let any agent act on a number, check it on your own data.

When it says 90%, it should be right about 90% of the time.

What About Other Open Models Like Clef?

Cloudflare released Clef on 1 October 2026 as open-weight decision models under Apache 2.0.

Cloudflare says they are Jev-API compatible and posted the weights on Hugging Face.

The main Clef model is built on a 27B base and the smaller Clef-flash is 9B.

I haven't tested Clef for business work yet, so I'm not going to make claims about its accuracy for you.

What I can say is that it's a far heavier model than Laya, so it'll need a lot more hardware to host yourself.

My Recommendation For Agencies And Small Teams

Start with a free hosted Jev door to prove the idea works on your real workflow.

Once you know which decisions you want automated, set up Laya locally and train it on your own examples.

Keep hosted Jev as the backup for big option lists and anything touching serious money.

That gives you speed, privacy and a cost of nothing per question, without betting the business on an untrained model.

If you want to turn this into a side income, my roundup of AI side hustles that actually work shows where this kind of setup work fits.

Also On Our Network

FAQ: Can I Run Jev Locally For My Business?

Can I run Jev locally on my own server?

No, Jev is only available as a hosted service and TypeSafe hasn't released its weights.

You can run Laya on your own server instead, which accepts the same request format and keeps client data in-house.

Is running a local Jev alternative cheaper than paying for Jev?

Per question, yes, because Laya costs nothing to run once it's installed.

But hosted Jev is already cheap at $0.042 per million input tokens, so the bigger reasons to go local are privacy and speed.

Is a local model accurate enough for sorting client leads?

Not straight out of the box.

In my test, Jev sorted 24 of 24 leads correctly and untrained Laya sorted 16 of 24, so train Laya on your own examples first.

Does local Jev keep client data private?

Running Laya on your own machine means the text you send it stays on that machine.

Hosted options, including the free ones, send your text to a third party, so read their data policy first.

Can I sell a local decision-layer setup to clients?

Laya's Apache 2.0 licence allows commercial use, so you can build it into client systems.

Just be honest with clients that it's a Jev-compatible alternative, not Jev itself.

About Julian

I'm Julian Goldie, an AI entrepreneur, SEO expert, and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation, and SEO.

I have 400,000+ YouTube subscribers who watch my AI tool tests every week.
I built a 7-figure agency from the ground up.
I run daily AI training inside the Boardroom.
I wrote two Amazon best-sellers on SEO and agency growth.

→ Get my best AI training inside the AI Profit Boardroom

So if a client asks "can I run Jev locally?", tell them no, then show them how a trained local Laya setup gives them the same kind of decisions for free while their data stays put.