A new AI paradigm: Jev

[chants]

Jev, released by TypeSafe AI in September 2026, is a "System One" decision model designed to return structured judgments rather than generate text. Unlike normal large language models (LLMs) which act as open-ended text generators, Jev acts as a classification and scoring engine built specifically to integrate with software logic.

https://typesafe.ai/blog/introducing-system-one-models-and-jev

The cost is way lower, it does not hallucinate, it is not auto regressive or token by token but gives back a predefined schema.

Looks extremely powerful and useful in a reversing workflow. Also hiding intent is far easier as it doesnt have all those intermediate thinking and output tokens to hit a filter. Looks important to the future of agentic reversing.

[chants]

Wow within days it is reverse engineered

https://archerhume.com/posts/jevs-architecture-unmasked/

Really interating writeup apparently they got it working on a MacBook

https://www.linkedin.com/pulse/whats-inside-jev-haixun-wang-lwxtc

[Fyyre]

Quoting from the below link: https://archerhume.com/posts/jevs-architecture-unmasked

"An ordinary LLM generates “90% confident” as text; its probability of producing those words does not establish a 90% probability of being right. Yet we build fraud screening, moderation, routing and risk assessment around precisely this pattern: paying for token-by-token generation, then treating an unvalidated confidence claim as a probability our software can act on."

That's extremely useful, and is something I have wished larger platforms provided. Thanks chants!

Originally Posted by chants

Jev, released by TypeSafe AI in September 2026, is a "System One" decision model designed to return structured judgments rather than generate text. Unlike normal large language models (LLMs) which act as open-ended text generators, Jev acts as a classification and scoring engine built specifically to integrate with software logic.



The cost is way lower, it does not hallucinate, it is not auto regressive or token by token but gives back a predefined schema.

Looks extremely powerful and useful in a reversing workflow. Also hiding intent is far easier as it doesnt have all those intermediate thinking and output tokens to hit a filter. Looks important to the future of agentic reversing.

[Shub-Nigurrath]

The concept is not new at all,
There is another engine which was published in March 2025 with a corresponding model. Unfortunately Laya is poorly trained, but the method used by Typesafe is exactly the same

https://arxiv.org/abs/2503.23303
https://huggingface.co/convaiinnovations/laya

I integrated it into my automatic RCE toolchain and indeed it give some advantages in the exploration phase, to rank the decision branches.

[Hypnz]

laya is far more better than jev....

[chants]

So you are right, Laya did come first in terms of "System One" models. It is also open weight.

Though HuggingFace own benchmark suggests Jev is better:

https://huggingface.co/blog/sora-2/jev-vs-laya-hosted-api-or-open-weights-2026-guide

74.4 vs 54.4 composite score. I think the big deal is Jev has 64k context while Laya only 512.

Of course benchmarks only tell a small part of a bigger story, so I certainly admit that Laya might be better beyond the fact it can be self hosted.

But there is something strangely satisfying in seeing Jev play Doom in real time: https://lumadock.com/blog/what-is-jev-typesafe