I asked Claude to help me whip up a Kotlin DSL and...

# feed

[pambrose]

I asked Claude to help me whip up a Kotlin DSL and client for Jev. Feedback would be appreciated.

https://github.com/pambrose/jev4k

[Stefan Oltmann]

I find Jev an interesting concept and wait for the first local model.

Given its performance I'm interested to see if that is fast and small enough to be bundled with an app (like we bundle face detection models, etc.).

By that time I'll check your lib out. 🙂

But please don't double post in the future.

[Didier Villevalois]

@Stefan Oltmann there are already plenty open-weight System 1 decision engines with RLCD. As a matter of fact, typesafe.ai kinda claims they invented this type of model, but as the creator of Laya explains in this writeup, he published two papers in 2025 about this...

Laya is really tiny, faster than Jev and more accurate in most benchmarks. And it is open-weight! 😁 You can get it here: https://huggingface.co/convaiinnovations/laya

There are also other open-weight System 1 models like those. You will find a lot by web-searching "open-weight System 1 decision engine with RLCD". To name two other than Laya: Qwen-RLCD, OpenJev (aka Verdict).

[Didier Villevalois]

@pambrose I would love for a Kotlin DSL and Kotlin engine that would work with those open-weight models! 😜

Laya multilingual has only 0.4B parameters and only weighs 644MB, so I guess that could run on small devices like phones...

[Stefan Oltmann]

Full ack!

Depending on a hosted service with API prices is not my thing, but I'd really like a ready-to-use KMP library that sets up everything to use Laya (runtime, etc.) offline/locally and offers an API like the one built for Jev here.

If @pambrose won't do it I might spent the next GLM Weekend Build tokens I can grab on it. 😬

[pambrose]

Given the Jev experience, we will likely see all sorts of competitors around the corner.

[Stefan Oltmann]

If the benchmarks can be trusted Laya is a lot better.
Maybe you can do your own benchmarks and confirm them?

[Stefan Oltmann]

I'm not interested at all in Jev cloud competitors or any cloud models.
Just in models small enough to run on consumer hardware that I can bundle with the app. Like I do with face detection and image classification

[pambrose]

Different strokes for different folks

[Stefan Oltmann]

What do you use Jev for?

[Stefan Oltmann]

Your lib has customer service examples. Is that your real use case?
Does it work like you expected?

[pambrose]

Classifiers are hardly something new and are applicable in all sorts of use cases. By themselves, they have limited value. When used in conjunction with an LLM, they can be very useful. I am using Jev in the context of a voice-enabled Koog project. It is a work in progress.

[Stefan Oltmann]

I'm looking forward your post when you have something to show. 😉

[Yassine Abou]

I don’t get the Jev hype. their marketing campaign was strong, but calling it 'X times faster' doesn’t mean much when it still loses to Qwen 3.6.

[Stefan Oltmann]

It loses to Qwen 3.6? How?

[Yassine Abou]

https://www.reddit.com/r/LocalLLaMA/s/Nj7wcc17L6

[Yassine Abou]

If you check the local Llama subreddit for JEV benchmarks, you’ll only find comparisons against Qwen 3.5. Why aren’t they confident enough to test against newer versions like 3.6, 3.7, or 3.8?

[Stefan Oltmann]

Interesting benchmark.

[pambrose]

@Didier Villevalois, your comments about Laya motivated me to dig into it a bit. I put together an implementation of the Jev API backed by Laya. It is not suitable for benchmarking or production use, but it is a cool way of playing with Laya from Kotlin.

https://github.com/pambrose/laya-server

[pambrose]

And yes, the ideal use case for Laya is probably as an embedded client, but using Python to run the model is much less complicated.

[Didier Villevalois]

@pambrose really nice! I'll give it a try.

Yeah, I think as a first try for running it locally, we'd have to transform the model + weights as an ONNX model, and an ONNX runtime.

[Stefan Oltmann]

I just found https://benchmarkheaven.com/jev-models

According to that Laya is not that great. :/

But it lists a lot of other options and ranks them.

[pambrose]

This is much better than the laya wrapper I wrote. It works with jev4k using these env vars:
TYPESAFE_API_KEY=whatever
TYPESAFE_DEFAULT_MODEL=laya
TYPESAFE_BASE_URL=http://localhost:11435

https://ollaya.dev

[Stefan Oltmann]

Or this: unsloth.ai/docs/models/decision-laya

[Didier Villevalois]

New kid on the block: CLM
https://contrastive-lm.notion.site/
https://github.com/Contrastive-LM/CLM

[Stefan Oltmann]

Very nice progress.

Now someone needs to turn something like that into an ONNX model that can be used from a normal app without hosting a vllm instance.