Evidence / Open source

Is Jev open source?

No. TypeSafe has not released Jev's weights or model code, and there is no public self-hosted build. Its customer agreement licenses Jev only as a TypeSafe-hosted API. OpenRouter, Vercel, and Cloudflare serve it too, but they host it for you.

The code around Jev is open: TypeSafe's JavaScript and Python SDKs, its coding-agent skill, and its System One adapter are all MIT-licensed. None of them contains the model. Several independent open-weight models now copy Jev's interface. The most prominent is Laya. They are separate models with their own accuracy and calibration, not a downloadable Jev.

Part	Status
Jev model weights	Not released
Jev model code	Not released
Self-hosted or on-prem Jev	No public option. Docs and terms describe hosted access only.
Architecture and training details	Described in outline (RLCD). No paper, parameter count, or training data.
API documentation	Public to read. It carries no open-source license.
SDKs, agent skill, System One adapter	Open source, MIT

Verified September 23, 2026, against TypeSafe's docs, its Master Customer Agreement, and its GitHub organization.

What TypeSafe has and hasn't released

Primary sources only

01Weights: not public. None of the ten public repositories in the typesafe-ai GitHub organization ↗ contains Jev weights. TypeSafe's site and docs link to no model hub. The models page ↗ says Jev isn't fine-tuned per customer and that the same weights serve every account through POST /v1/systemone. Two unverified Hugging Face accounts are named “TypeSafe AI.” TypeSafe doesn't link to either one, and neither holds Jev.

02License: hosted use only. The Master Customer Agreement ↗ (updated September 19, 2026) grants a limited license to “the TypeSafe-hosted” web console and API. There is no model license to download or redistribute. Section 2.3 bars customers from using the service or its output for model distillation, from training a model to imitate it, and from reverse engineering.

03Self-hosting: no public offering. TypeSafe's docs describe no on-prem, VPC, or offline deployment. The enterprise options they do mention are higher rate limits (via sales) and zero data retention. Whether TypeSafe would negotiate something private isn't public, so we don't claim it never would.

04Model code: not published. The official repositories are clients, instructions, and an LLM-backed adapter. The organization's vllm and LLaDA repositories are forks of public research projects, last updated in 2025. They are not a Jev release.

05Reproducibility: not enough published. The launch post ↗ names a new architecture, a parallel sampler, and RLCD training. It gives no paper, parameter count, dataset, or model card. Community write-ups that describe Jev's internals are inferences. See RLCD for what has been disclosed.

The official open-source code

All four repositories are MIT-licensed and published by TypeSafe. The two SDKs are clients for the hosted API. The skill is instructions for a coding agent. The adapter answers TypeSafe-shaped questions with a general LLM.

Repository	What it is	Contains Jev?	Works without TypeSafe?
typesafe-sdk-js ↗ MIT · v0.6.0	`@typesafe-ai/sdk`, the TypeScript client. Guide	No	Only against another server that speaks the same API. `baseURL` or `TYPESAFE_BASE_URL` points it there.
typesafe-sdk-python ↗ MIT · v0.7.1	`typesafe-sdk` on PyPI, the Python client	No	Same as JavaScript: set `base_url`.
skills ↗ MIT · v0.5.7	A Markdown skill that teaches Claude Code, Codex, and other agents to write TypeSafe code. Guide	No	Yes to install. It is instructions only and makes no calls.
system-one-adapter-python ↗ MIT · v0.2.1	A Python drop-in for the SDK's `system_one` call, backed by an LLM instead of Jev. TypeSafe built it to compare Jev with LLMs.	No	Yes. It calls OpenAI, Anthropic, Gemini, or any OpenAI-compatible endpoint you choose, and a local model server can be one.

Licenses read from each repository's LICENSE file and matched against npm and PyPI metadata. The Python SDK's LICENSE still has the MIT template's [year] [fullname] placeholders, but PyPI declares it MIT. The docs are public to read. That makes them documentation, not open source.

What “Jev-compatible” means

Projects use the phrase for very different things. We use four levels. Each includes the ones before it.

1Concept. Typed decisions from a single forward pass, with no generated text. The request and response format can be anything.

2Primitives. Choice, Score, and Noul (or clear equivalents), each returning probabilities.

3Wire. Implements POST /v1/systemone with TypeSafe's request and response JSON, so the same body works on either server.

4SDK. TypeSafe's own SDK runs unchanged apart from the base URL. We count this only when the SDK has actually been run against the project, by us or by the project.

None of the four levels says anything about behavior. A wire-compatible server can still disagree with Jev on your data, and its probabilities are calibrated differently or not at all. Changing the base URL swaps the model, so re-measure accuracy and reset thresholds.

Open models that do Jev's job

Checked against source repositories, September 23, 2026

These three publish trained decision weights and serve Jev's wire format. None is Jev, and none was trained or endorsed by TypeSafe. We included only projects we could check from the original repository or model card.

Project	Weights	Base and size	Runs on	Compatibility
Laya Convai Innovations	Public, Apache-2.0	ModernBERT-large encoder, 421M (English). mmBERT-base, 322M (multilingual).	CPU or CUDA GPU. A separate MLX port covers Apple silicon.	SDK. We ran both official SDKs against it (see below).
Kev ↗ jaredpalmer/kev	Public, Apache-2.0. LoRA adapters and decision heads, loaded on top of Qwen base models.	Qwen3.5 0.8B, 4B, 9B	CUDA, ROCm, Apple silicon (MLX). Project says 4B and 9B fit a 32 GB Mac.	SDK. The README calls it with TypeSafe's Python SDK. It also serves `/v1/models`. Not tested by us.
OpenJev ↗ openjev/openjev on Hugging Face	Public, CC BY-NC 4.0, so non-commercial use only. Helper code is Apache-2.0.	27B Qwen fine-tune. It reads option scores from a language model's first output position.	One 80 GB GPU with vLLM (FP8), or a Mac with the MLX 4-bit build (~15 GB)	Wire. The project says Jev clients can point at it. Not tested by us.

Kev follows a community reconstruction ↗ of Jev's architecture, which was inferred by probing the API. TypeSafe hasn't confirmed it. Kev's adapters load onto Qwen base models, which have their own licenses.

Servers that run an existing open model

These projects train no Jev-style model of their own. They ask a general open model for Jev-style answers by reading its option probabilities directly. Quality and calibration therefore depend on the base model you load.

Project	Code license	Model and hardware	Compatibility
razorback16/openjev ↗	Apache-2.0	NVIDIA's DiffusionGemma 26B-A4B checkpoint. Needs an NVIDIA GPU with 24 GB or more, or about 16 GB on Apple silicon.	SDK. The README shows TypeSafe's Python SDK pointed at it, and it accepts `jev-latest`. Not tested by us. It also promotes a hosted service (Codiv), which is not self-hosting.
Simple Jev ↗ Featherless AI	Apache-2.0	Any supported Hugging Face model, on CPU or GPU	Primitives, with a `/v1/systemone` alias. The project says it adds its own behavior and that its probabilities are not calibrated.
SemIf ↗ formerly “OpenJev”	MIT	Qwen3.5-4B and others, on one consumer GPU (the project uses an RTX 3090). llama.cpp and MLX backends also exist.	Concept. Its documented interface is its own CLI and JSONL format. It commits row-level outputs and its calibration method.

Laya

The open model most searches mean

Laya is published by Convai Innovations ↗. Its source lives at NandhaKishorM/laya ↗, maintained by Convai's CEO, and its Python package is laya on PyPI. It is an encoder with a trained decision head, not a generative model. Every question is answered in one forward pass. Two other popular repositories are independent runtimes for Laya's models: laya-mlx ↗ on Apple silicon and receptron/laya ↗ in Node.js. LayaAir, the game engine, is unrelated.

We tested the compatibility claim. In a scratch environment we ran laya-serve 0.3.7 locally and changed only the base URL. With that, TypeSafe's official JavaScript SDK (0.6.0) and Python SDK (0.7.1) both completed a Choice + Noul + Score request and parsed typed answers. Two differences showed up. GET /v1/models returns 404, so an SDK's model-listing call fails. Responses also carry extra routing and action fields. The SDKs tolerate them. Laya's server has no authentication unless LAYA_API_KEY is set, and it binds to all network interfaces by default.

!Its Jev comparison table combines separate studies. Laya didn't run Jev. It says so itself: it has no TypeSafe API access. Its Jev figures come from other studies with different samples and label sets. Its 0.766 typed-decisions result comes from a checkpoint fine-tuned on that benchmark's training split. The base checkpoints score near chance on that benchmark.

!Calibration needs fitting. The project reports that the checkpoints ship over-confident. Its better ECE numbers come after it fitted temperatures on held-out data. Do the same on your data before trusting a threshold.

=One independent head-to-head exists. Dhruv Mehra's September 22 benchmark ran Jev (via OpenRouter) and Laya's English checkpoint on the same 500 examples per dataset. Laya led on AG News (90.6% vs. 84.3%), a dataset in its own training mix, trailed on SST-2 (92.0% vs. 95.4%), and fell to 38.2% on 77-label Banking77, where Jev scored 76.4%. Latency isn't comparable, because Laya ran locally and Jev over the network. Third-party result, not reproduced by Jev FYI.

Laya's context is short: 512 tokens per question on the English checkpoint, 1,024 on multilingual, shared between state and options. Hosted Jev allows 32k for state plus the longest question. Long documents and choices with many options are its documented weak spots. Checkpoints, setup, limits, and the full Jev comparison are on the Laya page.

Which OpenJev?

No single project owns the name. These are the ones you're most likely to find, plus one popular project that dropped it.

Name and publisher	What it actually is
openjev/openjev ↗	The 27B fine-tune in the table above. Weights are non-commercial only. The project reports 84.0% against hosted Jev's 85.4% on its own 10,000-question set. It ran both, but the test is its own and hasn't been independently reproduced.
razorback16/openjev ↗	A server that runs NVIDIA's DiffusionGemma behind Jev's API. It publishes no weights of its own.
AlexWortega/openjev ↗	A Qwen3.5 model trained for entailment (NLI), MIT. It answers entailment, contradiction, or neutral, not Choice/Score/Noul, and it has no Jev wire API.
ZefanCai/Open-Jev-9B ↗	A LoRA adapter and decision head for Qwen3.5-9B, Apache-2.0, served on `/v1/systemone`. The project documents its training data. Its reported scores are on its own test splits.
TheoLeeCJ/SemIf ↗	Called OpenJev until it was renamed. Older articles that say “OpenJev” may mean this project.
OpenJEV (JEV) token	A cryptocurrency with no connection to TypeSafe or to any of the models here.

GitHub search returns more small repositories named OpenJev, many only days old. Pick by publisher, license, and weights, not by name.

Hosted Jev, an open model, or neither

People searching “open source” usually need something concrete: data kept in-house, offline operation, a model they can fine-tune, or no vendor dependency. Match the option to the actual need.

If you need	Realistic option
The real Jev	Hosted only: TypeSafe, OpenRouter, Vercel AI Gateway, or Cloudflare Workers AI. See API & access and pricing.
Less data retained by the vendor	Before self-hosting, check whether TypeSafe's zero data retention for enterprise customers is enough. Data still leaves your network either way.
Data never leaves your hardware, or offline use	An open model: Laya or Kev on modest hardware, or OpenJev if you have a large GPU and a non-commercial use. You operate, patch, and monitor it yourself.
To keep existing TypeSafe SDK code	Choose a wire- or SDK-compatible server and change the base URL. Then re-run your evaluation set, because thresholds tuned on Jev won't carry over.
A model trained on your own labels	Laya and Kev publish training code. Jev can't be fine-tuned. The customer agreement bars training a model to imitate Jev's output, so a local copy distilled from Jev is off the table. TypeSafe does document a different pattern: using Jev's answers as features ↗ for a model trained on your own labels.
Just local typed classification	You may not need a Jev clone at all. With labeled data, a small fine-tuned classifier beat Jev on two of three public datasets in an independent benchmark. See classification for when rules or embeddings fit better.

−What you give up by self-hosting. Hosted Jev accepts 32k tokens of state plus the longest question. Laya handles 512 to 1,024 tokens per question. You also give up TypeSafe's trained calibration, and every alternative here reports needing its own temperature fitting. Every open alternative on this page appeared in the past two weeks, and most are maintained by one person or a small team. Multilingual quality varies. Jev's docs say English is its strongest language, and Laya routes other languages to a separate checkpoint.

+What you gain. Weights that change only when you change them, no per-token fees, no rate limits, and no network round trip. The costs move to your own hardware and staff time. We haven't compared total cost.

What we left out, and why

×Von (wfzyx/von, Apache-2.0, ModernBERT-based). Its README states Jev's architecture as fact (“proprietary MoE”), but TypeSafe hasn't disclosed its architecture. It also presents its calibration as guaranteed. We left it out until its claims are narrower.

×openJev-verdict-2.0. Its headline claim, that it beats Jev and Laya, comes from its own run on a single benchmark. Its Jev row is copied from Laya's published table, and it gives Jev a parameter count that TypeSafe has never published.

×NanoJev (TianyuCodings/NanoJev). A well-documented 0.6B research replica trained for game tasks, not general decisions. Its model card states no weights license.

×Directory counts. We don't repeat the “35 alternatives” style of tally. Many listed projects are demos, forks, or thin wrappers, and a count says nothing about quality.

Every number on this page is attributed to whoever measured it. Jev FYI has not benchmarked any of these models. For Jev's own evidence, see benchmarks.

← Use hosted Jev TypeSafe & the claims ledger →Jev benchmark evidence →