Is Jev Open Source? Weights, Paper and Running It Locally
Last checked · Independent guide, not affiliated with TypeSafe AI
No. Jev is a closed model served only through TypeSafe's API and resellers. TypeSafe has not released weights, a technical paper or a way to run it locally as of September 19, 2026. What it does open-source are its Python and JavaScript SDKs, an agent skill, and an adapter that runs the same questions on OpenAI or Anthropic models.
Open weights were one of the first things people asked for after launch. “Any open weights models that can do this yet?” was asked in the Hacker News launch thread, and searches for “jev github” rose alongside the main product name.
What TypeSafe has and has not released
Section titled “What TypeSafe has and has not released”| Item | Status (Sep 19, 2026) |
|---|---|
| Model weights | Not released |
| Technical report or paper | Not published |
| Architecture details | High-level only (see Is Jev an LLM?) |
| Self-hosted or on-premises version | Not offered publicly; TypeSafe’s docs point enterprise buyers to sales for custom plans |
| Python SDK | Open source on GitHub (typesafe-sdk on PyPI) |
| JavaScript SDK | Open source on GitHub (@typesafe-ai/sdk on npm) |
| Agent skill for Claude Code, Codex and others | Open source on GitHub |
| System One adapter (same questions answered by an LLM) | Open source on GitHub |
The SDKs are clients: they call TypeSafe’s hosted API and contain no model. Installing typesafe-sdk does not let you run Jev offline.
Can I run Jev locally?
Section titled “Can I run Jev locally?”No. Every call goes to a hosted service: TypeSafe’s API, or OpenRouter, Vercel AI Gateway or Cloudflare Workers AI, which pass requests to the same model. TypeSafe also states that Jev is not fine-tuned per customer; everyone uses the same weights.
If you need something on your own hardware, the options are community projects that reproduce Jev’s interface on small open models. SemIf (formerly OpenJev) is the best known and even runs in the browser. These projects are explicit that they do not reproduce Jev’s model or training. See Jev alternatives.
What about the “open-source Jev” posts?
Section titled “What about the “open-source Jev” posts?”In Jev’s first week, several posts on Hacker News and Reddit claimed to have open-sourced Jev or even beaten it. On inspection they fall into three groups:
- Interface copies that read option probabilities from an existing open model (SemIf, mini-jev).
- New small models trained for the same kind of task (jevlike).
- Existing classifiers reposted under Jev’s name, such as GLiClass, which dates from 2024.
None of them is Jev, and their benchmark claims have not been independently reproduced.
Is there a Jev paper?
Section titled “Is there a Jev paper?”Not yet. TypeSafe’s launch post includes evaluation charts and a description of its training method, RLCD, but no paper, model card with training details, or calibration study. Commenters on Hacker News asked for a downloadable technical report; as of September 19, 2026, none has been published.
Why it matters
Section titled “Why it matters”A closed model means your application depends on TypeSafe’s service: its status, its rate limits and its pricing. Two habits reduce the risk: pin the model version (jev-1.13.0) so answers do not shift under you, and keep your questions and thresholds in one place so you can re-test them against a different backend, which is exactly what TypeSafe’s own adapter is for.