No, the Jev model is not open source. You can’t download its weights or see how it was trained; you reach it through a hosted API. But most of what you touch around Jev is open: the official SDKs, TypeSafe’s agent skill, and the community MCP servers are MIT-licensed, and there are open-weight models that answer the same kinds of questions on your own hardware.
The short answer
People search “Jev open source” for two different reasons: to find out whether they can run Jev themselves, or to find open code to build with. The answers differ.
- Run it yourself? No. Jev is a hosted model. TypeSafe’s documentation says Jev “is not fine-tuned or LoRA-adapted with customer data” and that the same weights serve every account. You call it over an API.
- Build with open code? Yes. Everything from the SDKs to agent integrations is on GitHub under permissive licences.
What’s open and what isn’t
“Open source” usually means you can read, change, and redistribute the code. For AI models people also ask about open weights: whether the trained model file is published. Jev has neither, but that only covers the top of the stack.
Open source from TypeSafe
TypeSafe AI publishes these under the MIT licence, which lets you use, change, and ship them in commercial products:
| Project | What it is | Install |
|---|---|---|
| typesafe-sdk-python | Official Python client, sync and async, with retries | pip install typesafe-sdk |
| typesafe-sdk-js | Official TypeScript/JavaScript client for Node.js 20+ | npm install @typesafe-ai/sdk |
| skills | Agent skill that teaches Claude Code, Codex and others to write Jev code | claude plugin install typesafe@typesafe-ai |
| system-one-adapter-python | Same client interface, answered by an ordinary LLM API instead of Jev | See README |
The adapter is worth a second look. Because it keeps the same interface, you can run your questions through Jev and through an LLM with the same code and compare them, which is a cheap way to check whether Jev is right for your task.
Open-source community projects
Within days of launch, developers had published dozens of projects on top of Jev. The most useful groups:
- MCP servers that let coding agents call Jev, such as the System One Connector and JevMCP (both MIT). Our Jev MCP guide compares them.
- Integrations such as the LiteLLM pass-through and Spring AI support.
- Demos and tools: code review triage, Postgres filtering, game agents, music theory, and more.
We keep a checked list, with licences and caveats, in the Jev ecosystem directory.
Open-weight alternatives to Jev
If you need the model itself to be open, these projects answer the same three kinds of questions (yes/no, choice, score) and can run on your own hardware. None is a copy of Jev; each was trained independently.
| Model | Licence | Size | Reads per question | Notes |
|---|---|---|---|---|
| Laya | Apache-2.0 | 421M (English), 322M (multilingual) | 512–1,024 tokens (up to 8,192 multilingual) | pip install laya. Runs on a laptop. Server speaks Jev’s API format. |
| Open-Jev | MIT (adapters) | 2B, 9B, 27B | Depends on base model | LoRA adapters on Qwen; base weights downloaded separately. Needs a GPU for the larger sizes. |
| CLM-v0.1-8B | See model card | 8B | 2,048 tokens by default | clm-serve implements /v1/systemone. |
Compare that with Jev’s 32,000 tokens of state per request and up to 255 options per choice. The open models are much smaller, which makes them fast and cheap to run, but they see less text and handle long option lists less well. We tested Jev and Laya on the same data in Jev vs Laya.
Hosted Jev or an open model?
A practical middle path: start with hosted Jev to find questions that work, then test an open model on the same labelled examples if privacy or cost pushes you toward self-hosting.
Can you self-host Jev?
No. But you can write code that works with both. Laya’s server and CLM’s server accept the same POST /v1/systemone request as TypeSafe’s API, so switching can be as small as changing the base URL. The System One Connector MCP server supports this directly:
# Point an MCP server at a Laya server on your own machine
TYPESAFE_API_KEY=local TYPESAFE_BASE_URL=http://127.0.0.1:8787 evaluate setup mcpAnswers won’t be identical between models, so re-check your thresholds after switching. For how to design questions that transfer well, see how to use Jev.
Frequently asked questions
Is Jev open source?
No. Jev's model weights and training method (RLCD) are proprietary to TypeSafe AI, and the model is only available through hosted APIs. TypeSafe's SDKs, agent skill and adapter library are open source under the MIT licence.
Can I download the Jev model weights?
No. TypeSafe does not publish Jev's weights, and the same weights serve every customer. There is also no customer fine-tuning; you shape Jev's answers through the state, instructions and criteria you send.
Is there an open-source version of Jev?
There is no open version of Jev itself, but several open models answer the same kinds of typed questions. Laya (Apache-2.0) is a 421M-parameter encoder you can install with pip; Open-Jev (MIT) adds adapters to Qwen models; CLM-v0.1-8B is another open System One model. None of them is a copy of Jev.
Can I run Jev locally or offline?
No. Jev is hosted by TypeSafe and resold through Vercel AI Gateway, OpenRouter and Cloudflare. For offline or on-premises use, run an open model such as Laya; its server speaks the same /v1/systemone format, so code written for Jev can often point at it.
What licence do the Jev SDKs use?
The official TypeSafe Python SDK (typesafe-sdk-python), JavaScript SDK (typesafe-sdk-js), agent skill and System One adapter are all MIT-licensed on GitHub, as of 26 September 2026.
Sources
- TypeSafe docs Models: customising Jev and data handling ↗
- GitHub typesafe-ai organisation (SDKs, skills, adapter) ↗
- GitHub NandhaKishorM/laya ↗
- Hugging Face convaiinnovations/laya ↗
- GitHub zefan-cai/open-jev ↗
- Hugging Face Contrastive-LM/CLM-v0.1-8B ↗
- GitHub itsmostafa/typesafe-mcp (System One Connector) ↗
Licences were read from each repository’s licence file on 26 September 2026; they can change, so check before you depend on one. This site is not affiliated with TypeSafe AI or any project listed.