The Jev model is hosted
Plan around an API. The documented endpoint is POST https://api.typesafe.ai/v1/systemone, authenticated with a TypeSafe key. There is no self-hosted Jev binary and no Hugging Face weight drop as of 20 September 2026.
That matters for procurement, data residency, and incident response. If Jev is unavailable, your branch still needs a fallback — rules, a classifier, or a larger model.
Open tooling is not an open model
Examples, MCP wrappers, LangChain tools, and awesome-jev lists can be public while the model stays closed. Check the current TypeSafe terms before you promise a customer that Jev is open source, on-prem, or air-gapped.
The useful questions for a production review: where does state go, how long is it retained, can you pin jev-1.13.0, and what happens when jev-latest moves.
If you actually need local inference
Jev is a fit when latency, typed output, and a hosted API beat model ownership. If you must run on your own GPUs, look at open-weight classifiers, embedding plus rules, or a small local LLM with a strict schema. Those systems will not match TypeSafe’s published Jev numbers, and you should not advertise them as Jev.
- Jev: hosted, typed decisions, no public weights.
- Open reproductions: independent experiments, different models.
- Local requirement: pick an open stack and keep the names honest.
FAQ
Is Jev open source?
No. Jev is a closed managed API from TypeSafe AI. The model weights, architecture, and RLCD training details have not been released.
Can I download Jev weights or run Jev locally?
Not from TypeSafe. If local inference is a hard requirement, compare open classifiers, small instruction-tuned models, or community projects that only mimic the interface.
Are the Jev SDKs open source?
The client libraries (typesafe-sdk for Python and @typesafe-ai/sdk for Node) are installable packages around a hosted endpoint. That is not the same as releasing the model.
What about openjev or jevlike?
Community projects appeared within days of the launch. They explore similar typed-decision interfaces on open models. They do not reproduce TypeSafe’s model or training method.