No. TypeSafe has not published Jev’s weights, parameter count, architecture or training recipe. Its SDKs and examples on GitHub are open source under MIT, but they are clients for the hosted API, not the model.

Can I download Jev or run it locally?

No. There is no official checkpoint to download, on GitHub, Hugging Face or anywhere else. Jev is only available as a hosted API, from TypeSafe directly or through services such as JevStation. If your data cannot leave your machines, you need an open-weight model instead.

Is there a Jev model on Hugging Face?

Not from TypeSafe. As of 2026-11-05 TypeSafe has no models on Hugging Face. Models such as autotrust/JEV-35B or JevK5 use the name, but their own model cards say they are independent projects, built on Qwen base models, and not affiliated with TypeSafe.

What is the official Jev GitHub?

TypeSafe’s organisation is github.com/typesafe-ai. Its main repositories are the Python SDK (typesafe-sdk-python), the TypeScript SDK (typesafe-sdk-js), agent skills (skills), an adapter that mimics the client with LLM APIs, an n8n node and public examples. None of them contain Jev’s weights.

Is there a Jev paper or architecture description?

No full paper. TypeSafe describes Jev as a System One model trained with RLCD on synthetic data, but has not published the architecture, parameter count or calibration curves. Outside guesses that it is built on an open-weight LLM are unconfirmed.

What can I self-host instead?

Cloudflare’s Clef and Clef-flash are open-weight decision models under Apache 2.0 that accept Jev’s request format. Community models such as JevK5 and SemIf copy the interface on small open models. None of them is Jev, and most have only been benchmarked by their own authors.

Is Jev open source? What is on GitHub and Hugging Face

No. Jev’s weights, architecture and training recipe are not public, and there is nothing to download: Jev runs only as a hosted API. What TypeSafe does publish on GitHub is code around the model — MIT-licensed SDKs for Python and TypeScript, agent skills and examples. The models named “JEV” on Hugging Face are made by other people and are not TypeSafe’s Jev.

JevStation is an independent workspace for Jev and other decision models. It is not TypeSafe’s official site, and the repositories below are listed for reference, not because we run them.

更新于 2026-11-05

The short answer

What people look for	Exists?	Who publishes it	Open source?
Jev’s weights / a download	No	—	—
A Jev paper or architecture	No full paper	—	—
Official SDKs and examples	Yes	TypeSafe, on GitHub	Yes, MIT
“JEV” models on Hugging Face	Yes, but they are not Jev	AutoTrust AI and others	Yes, mostly Apache 2.0
Projects that call Jev	Yes, many	Independent developers	Mostly MIT; they need an API key
Open-weight models like Jev	Yes	Cloudflare, the community	Yes

So “Jev open source” has two honest answers. The model is closed. The code around it, and a growing set of look-alike models, is open.

What TypeSafe publishes on GitHub

TypeSafe’s organisation is github.com/typesafe-ai. These are its public repositories that matter if you are building with Jev, as of 2026-11-05:

Repository	What it is	Licence
typesafe-sdk-python	The official Python client, `pip install typesafe-sdk`	MIT
typesafe-sdk-js	The official TypeScript/JavaScript client, `npm install @typesafe-ai/sdk`	MIT
skills	Agent skills for building with the System One API	MIT
system-one-adapter-python	A drop-in replacement for the Python client that calls LLM APIs instead	MIT
n8n-nodes-typesafe-ai	An n8n node for the API	—
typesafe-public-examples	Content used in the docs, demos and cookbooks	—

All of them are clients, tooling or examples. Every one of them needs TypeSafe’s hosted API, or an LLM in the adapter’s case, to produce an answer. There is no repository with Jev’s weights, and none named after the model.

Not everything with “TypeSafe” in the name is TypeSafe’s. A separate GitHub organisation called “TypeSafe Community” was created on 2026-09-18 and links to a different website; we could not confirm it belongs to the company. Package registries have the same problem: jev-lite on PyPI calls itself a client for Jev but is not the official SDK, which is typesafe-sdk. Check the publisher before you install anything that will hold your API key.

The “JEV” models on Hugging Face are not Jev

Search Hugging Face for “jev” and you get about a hundred models, some with over a million downloads. None of them is from TypeSafe, which has no models on Hugging Face at all.

autotrust/JEV-35B, JEV-27B, JEV-9B and variants. Open-weight decision models by AutoTrust AI, fine-tuned from Qwen base models, under Apache 2.0. Their model card says plainly that they are independent and not affiliated with, endorsed by or a product of TypeSafe. The benchmark figures on the card are the authors’ own.
JevK5. A 4B model on Qwen3.5-4B with a self-hosting runtime (allebee/jevk5), described by its author as an “open-weight Jev alternative” and not Jev’s model or architecture.
Smaller “Jev-style” models. A number of 0.8B–2B checkpoints for single jobs, such as mail or dependency checks, that copy Jev’s answer format.

These can be useful, but downloading one does not get you Jev. They are trained by different people, on different data, and answer differently. Treat their numbers as vendor claims until someone else reproduces them.

Why there is no Jev paper or architecture

TypeSafe describes Jev as a System One model: it reads a state and typed questions and returns a probability for every option instead of text. It says the model is trained with RLCD (Reinforcement Learning for Calibrated Decisions) on synthetic data and that the same weights serve every account. It has not published the parameter count, the architecture, the full training recipe or calibration curves. Observers have suggested it is built on an open-weight LLM; TypeSafe has not confirmed that.

What a System One model is covers what is known in more detail.

Open-source projects built on Jev

These repositories are open source, but they call the hosted Jev API, so you still need an API key to run them. We have written a setup guide for each:

Project	What it does	Our guide
browser-use/jev-ultrafast	A fast, cheap web agent	jev-ultrafast
AgriciDaniel/jev-seo	A live SEO audit of a website, judged by Jev	jev-seo
devagrawal09/jev-review	A staged code-review workflow with a local dashboard	jev-review
gargpratyush/jev-router	Routes Claude Code and Codex to the cheapest suitable model	jev-router
sharziki/semdecide	Typed semantic decisions for shell pipelines and CI	semdecide

None of these projects is affiliated with TypeSafe or with JevStation. A JevStation API key works with any project that lets you set the endpoint; new accounts get 50 free credits, no card (how to get Jev).

If you need something you can host yourself

You need an open-weight model when the data cannot leave your infrastructure, when you must run offline, or when you want to fine-tune on your own labels. The options, in short:

Clef and Clef-flash (Cloudflare). 27B and 9B decision models under Apache 2.0 that accept Jev’s request format. The closest thing to an open Jev, though its benchmarks so far are Cloudflare’s own. See Jev vs Clef.
Community look-alikes. JevK5, SemIf, open-jev and jev-on-a-laptop copy the interface on small open models. The one that measured itself against Jev scored about 13 points lower.
A classifier you train yourself. Best when the label set is small and stable and you already have plenty of labelled data.

Jev vs open models compares these with the measurements we could find. To check one against hosted Jev on your own input, the playground runs Jev, Clef and Clef-flash side by side.

Sources

TypeSafe repositories and licences. github.com/typesafe-ai, checked 2026-11-05 (vendor)
Package names. typesafe-sdk on PyPI and @typesafe-ai/sdk on npm (vendor)
No TypeSafe models on Hugging Face; third-party “JEV” models and their disclaimers. Hugging Face model search and the autotrust/JEV-35B and JevK5 model cards, checked 2026-11-05 (independent)
What TypeSafe has and has not disclosed about the model. What a System One model is (our summary)

常见问题

Is Jev open source?: No. TypeSafe has not published Jev’s weights, parameter count, architecture or training recipe. Its SDKs and examples on GitHub are open source under MIT, but they are clients for the hosted API, not the model.
Can I download Jev or run it locally?: No. There is no official checkpoint to download, on GitHub, Hugging Face or anywhere else. Jev is only available as a hosted API, from TypeSafe directly or through services such as JevStation. If your data cannot leave your machines, you need an open-weight model instead.
Is there a Jev model on Hugging Face?: Not from TypeSafe. As of 2026-11-05 TypeSafe has no models on Hugging Face. Models such as autotrust/JEV-35B or JevK5 use the name, but their own model cards say they are independent projects, built on Qwen base models, and not affiliated with TypeSafe.
What is the official Jev GitHub?: TypeSafe’s organisation is github.com/typesafe-ai. Its main repositories are the Python SDK (typesafe-sdk-python), the TypeScript SDK (typesafe-sdk-js), agent skills (skills), an adapter that mimics the client with LLM APIs, an n8n node and public examples. None of them contain Jev’s weights.
Is there a Jev paper or architecture description?: No full paper. TypeSafe describes Jev as a System One model trained with RLCD on synthetic data, but has not published the architecture, parameter count or calibration curves. Outside guesses that it is built on an open-weight LLM are unconfirmed.
What can I self-host instead?: Cloudflare’s Clef and Clef-flash are open-weight decision models under Apache 2.0 that accept Jev’s request format. Community models such as JevK5 and SemIf copy the interface on small open models. None of them is Jev, and most have only been benchmarked by their own authors.