TypeSafe AI’s Jev does not write. It answers. Released on September 15 and covered in detail by ITmedia, the model takes a question you define - classify this, choose between these, score that, is this true - and returns a typed answer as structured data with a confidence value. The company calls the category a System One model, and the interest is easy to explain: most of what an agent asks a frontier model to do is not writing, it is choosing.
Key facts
| Fact | Detail |
|---|---|
| Question answered | Jev and the System One Split for Agent Decisions |
| Short answer | TypeSafe AI’s Jev does not write text - it returns typed decisions as JSON, and Vercel, Cloudflare and others have already wired it in. |
| Reading time | ~2 min |
| Last updated | 2026-09-21 |
The fast/slow split, stated plainly
QbitAI’s write-up puts the architecture in one line: expensive models handle the slow thinking (planning, exploring alternatives, recovering from failure), while high-frequency atomic fast judgements - classification, selection, scoring - go to a lightweight decision model.
Two implementation details from that piece are worth keeping:
- The point is to skip autoregressive decoding and score directly, which is what makes the answer cheap and immediate rather than a short essay.
- The engineering layer that assembles and routes the two halves has a name now: the harness. That is where the split is actually enforced.
Adoption is the part that makes this more than a demo: per Forbes via Techmeme, Vercel, Cloudflare and others have already added Jev, with the claim that it matches much larger models on workflow evaluations. China’s APUS published one of the first independent open-source reimplementations, packaged as an agent skill that runs offline on macOS, Linux and Windows - including machines without a GPU.
Why the split matters on your own machine
A two-tier agent is cheaper to run all day, and that is the whole argument for it. But notice what the fast half is allowed to decide: which option, not whether the action should happen. Those are different questions, and only the second one carries risk.
That distinction is exactly where a desktop agent earns or loses trust. On MOOGH the split is explicit: it plans with the slower reasoning, shows you the proposed file and command changes, waits for your approval, and records what it did. A faster model underneath makes the work quicker; it does not move the decision about permission. Download MOOGH if you want the fast half fast and the final call yours.
Related reading: