Week of September 14, 2026

Jev Decision Index

Within a week there were 31 open reproductions of Jev, and multimodalart’s Decision Index scores every one of them on the same frozen 132,422-decision suite Jev itself was run on. Jev still leads at 59.5; the best open entry, a Qwen3.8-27B setup called Jevfire, is at 55.7, with a diffusion-Gemma 26B right behind it. Where Jev pulls away is knowledge and reasoning — 78.3% on GPQA Diamond against 48.0% for the best open model — while the open models already match or beat it on language understanding and retrieval.

So the alternatives are real, and some are small enough to run yourself — which also sidesteps the licence questions below. I still think Jev is the one to use for now: it’s the strongest, it’s cheap, the API is convenient and they host it for you. (Numbers are from the index’s Sep 22 update; Jev’s latency there is a hosted round trip, so don’t read anything into the speed comparison.)

No moat, and Jevons says that’s fine

By the end of the week the backlash had arrived. Tim Kellogg called Jev overhyped with no easy moat — every lab will have an equivalent in a few months — and someone surfaced Laya, whose author says they built non-autoregressive decision models a year earlier. I can’t see a moat either : once the idea and the use cases are out there, there are dozens of ways to build something similar, and OpenAI’s and Anthropic’s GPUs can eat Jev’s lunch before breakfast. The easiest path to a Claude-Jev might just be an acquisition, if TypeSafe doesn’t get too greedy on price.

But I don’t much care whether it’s Jev or Jev-like — the concept is the important part, and Jev is what got people to notice. And the name is apt: Jevons’ paradox says Jevving the Jevvable problems probably increases demand for regular LLMs rather than eating into it. (For the record, Laya’s page says it outperforms Jev; on the Decision Index above it scores 16.4 to Jev’s 59.5.)

Read the licence before you Jev at work

Tim spotted that TypeSafe’s terms made Jev personal-use-only unless you asked permission, and when I read the rest it was kinda a shitshow for any professional use: getting acquired could void your contract, no security testing but unlimited liability, no using the output to build anything “similar or competing” (undefined), terms that were themselves marked confidential, and a clause saying output “similar or identical” to another customer’s isn’t yours.

To their credit, TypeSafe moved fast. Tim reported they updated the terms on Sep 19 — no more permission needed to build products on it, the MCA-vs-ToS conflicts cleaned up, no benchmarking ban. Not everything went, though: the current Master Customer Agreement still bans security testing, “similar or competing” products, and still says similar or identical output isn’t yours. Fine for side projects; get legal to read it before production.

What Jev is, and what you’d use it for

TypeSafe launched Jev on Sep 15, and it doesn’t generate text at all. You hand it arbitrary input plus the outputs you want, and it hands back all of them at once, typed, each with a calibrated probability — basically a very smart multi-output classifier with a powerful encoder. It does continuous values too (sales prediction, churn risk, anomaly detection), and it’s cheap: $0.042 per million input tokens, output free, answers in 70–500ms.

So what’s it for? All the usual classifier stuff — robotics, workflow routing, decision support — except the input is arbitrary text through a smart encoder. The one I like most is routing : hand Jev the query, have it score the possible model/effort combos, and send the work to the right full LLM. More generally, a lot of apps are “get the model to say something” followed immediately by “parse it” (as catblanketflower put it), and anything shaped like that is Jevvable.