Short answer: no. Jev is a proprietary model served by TypeSafe AI through its API. The weights are not published, there is no self-hosting option, and the model cannot be run offline. Anyone telling you otherwise is selling something.

What is closed

  • The weights. No checkpoint download, no quantized community build, no run-it-on-your-laptop variant.
  • The serving stack. Every call goes to TypeSafe’s endpoint at api.typesafe.ai. When it is down or slow, your decision path is down or slow.
  • The training recipe. TypeSafe has described its RLCD training approach publicly at a high level — its founder’s launch post discusses it — but the data, code, and process are not open.

This matters for some architectures more than others. If you need air-gapped deployment, data-residency guarantees, or to freeze a model version forever, Jev is not currently an option, and no amount of wrapper tooling changes that.

What is open

The ecosystem around the model is a different story, and it is unusually open for a three-week-old model. Browse the GitHub entries in this directory and the pattern is immediate: SDKs and clients in several languages, MCP servers exposing Jev as a tool, agent harnesses that route through it, evaluation suites that score it against alternatives, and end-to-end applications published with their source. Builders ship Jev-adjacent work in the open because the interesting part — the harness, the criteria, the loop — is theirs, not the vendor’s.

A useful way to see it: Jev itself is closed like a database engine is closed, while the schemas and queries built on it are open. The decision criteria you write, the state assembly, the confidence routing — all of that is your IP and most of it publishes well.

The questions that actually matter

Openness questions tend to hide the questions builders should ask first:

1. What does it cost to leave? With Jev, the closed part is one HTTP call with a documented request format. Your criteria and state assembly are portable; swapping the endpoint for another decision model — or a prompted fallback behind the same interface — is an adapter, not a rewrite.

2. Is the interface documented and stable? The request and response format is public. The vendor can change pricing or deprecate versions; pin what you can and watch changelogs, same as any managed dependency.

3. Does the vendor lock the data? Judgments are state in, answers out. There is no fine-tuning loop that entangles your data with the model.

If open weights are a hard requirement

Then Jev is out for the core call, and the honest option is a prompted LLM behind the same interface your code already speaks — with the tradeoffs laid out in Jev vs an LLM: you regain control and lose the typed guarantees, the speed, and the price. Some builders run both, one primary and one fallback, precisely because the wrapper makes them interchangeable.

For everyone else, the practical reading is: the thing you build on Jev is yours and can be open; the model is a metered utility. The receipts in this directory — cost, latency, source links — are the work of people treating it exactly that way.

Published 25 September 2026. Status of weights, hosting, and API availability checked against TypeSafe AI’s public documentation on this date.