Clef: Open-source decision models, and new RL fine-tuning platform (blog.cloudflare.com)

It's not a "new paradigm", it's a low-hanging fruit that's been lying around for years; Typesafe were the first to bother to stop and pick it up, and market the shit out of it. But it was still a low-hanging fruit.

There are many, many of those left around, because AI frontier is moving forward so fast, everyone is racing ahead. Which is why I laugh when people say AI is not transformative and LLMs are a dead end (and my favorite, "what are we going to do with all those GPUs when the bubble pops?"). Even if SOTA LLMs hit a hard capability limit tomorrow and never advanced again, there's a good decade of growth and advancement to be extracted just from all the low-hanging fruits that were left unpicked along the way.

They have to have decision models already in production. Part of their business is detecting bots, DDoSers and spammers.

A lot of people claim to have made better than jev, there's a jev benchmark, I have tried many of those models and they eventually end up failing, a non trivial task which doesn't seem like much but reminds me of the svg pelican bench is games, have one of these decision/classifier models play a game, hook it up to the input, most of the ones that are supposedly on jev level end up playing a terrible game, showing that they are very narrow. Cloudflare doesn't compare to the top open bench alternatives, I just finished downloading it and will compare it to jev for non trivial tasks tonight.

Output for decisions have so few output items (not really tokens here) that they are negligible anyway. Jev hyping "free output" is almost lying by omission.

> no output price listed

It's weird to think of these kinds of models as having "output tokens". Cross-encoder approaches like Laya add a [MASK] marker per option, but nothing is generated the way an autoregressive transformer generates. It's one bidirectional pass over your input, then a small head scores each option, so you wouldn't really pay for output as much as only input

Came directly to comments hoping not to see this one.

<sad trombone sound>

Surely someone will soon do what the title of this post makes it seem like cloudfare did. Truly modular open source training and inference logic, along with a totally open corpus and weights, will eventually out-compete the closed ecosystem.

Yeah I also thought it was strange their pareto frontier didn't include cost.

> and Clef-flash is based on Qwen3.8-9B

There is no official qwen 3.8 9b

From the model card:

> Clef-Flash is post-trained from Qwen/Qwen3.5-9B. See Clef for the larger variant.

Atom is 60M Param (around 133x to 400x smaller).

16ms latency. And locally run.

https://at0m.pienomial.com/

Why go big when you can go small ?

tbh saying all these dumb decision models are similar to Jev is like saying markov chains weren't far from GPT-2.

The value isn't really in the I/O shape, it is in the intelligence combined with the output shape. Every extra ounce of intelligence in these models unlocks additional use-cases. But the converse is also true: a dumb decision model is going to be less useful than using a more intelligent standard LLM.

That is the appeal of Jev: for certain usage it has more intelligence than some small SOTA LLMs. It is the first decision model that actually feels intelligent (to me).

I think better than priors would be a closed loop where you tell it what the right answer was (or some signal) and they monitor and fine-tune for you

We were able to completely automate 20,100 token prompts with At0m[https://at0m.pienomial.com/].

We believe entire compliance workflows (even multilingual) could be automated.

Would you like to get a demo ?

Also one of the more interesting features of Jev is the confidence rating that hardly any Jev-cc talks about.

It seems like jev's major advantage over existing classifiers is that I dont have train it.

If I have to gather and tag data to fine-tune Jev, I can probably just train an "old school" classifier model and make it even cheaper, faster, and just as accurate.

I suppose a sort of prior-proxy can be encapsulated by a carefully written system prompt.

You can make a basic one in minutes based on existing open-source models.

Latency won't be that good, but could still work similarly. Simply force the structured output of a LLM to the given schema.

Probably also easy to train because we can use stronget LLMs to generate input/output data, or even synthetic data is easy to generate.

It's not really a new technology, it's more like a new use-case.

Yes it's very easy if you have fairly basic ML knowledge.

Live coding Jev from Scratch | Understanding Qwen architecture

https://www.youtube.com/watch?v=AzxoU7kxjig

The basics are pretty simple. And depending on what your specific need is, the model can be really really basic, fast and super effective (ie. run on a mobile device and process thousands of requests in <100ms)

I've been playing with this for the last year or so. Started with a personal email classifier, also did benchmarks with some public datasets, then created a couple classifiers that could play Doom, and now I've been trying out some other experiments, like a request proxy/router to automatically choose a classifier and fallback to LLM to handle unseen requests

Jev did a great job at creating hype, but also at shaping the concept and space of "decision engine" or "decision model". People were already doing this with LLMs, which is very inefficient for most tasks like that, and the Jev guys figured there was a market there. It seems like they were right, and now there's a rush to flood the space, taking advantage of the hype window

There is a bunch of stuff to tease apart.

In general, training a general purpose classifier is something lots of people have worked on for a long time. Large Transformer models themselves are typically "generalists" already, so structured generation and constrained decoding have given you the ability to use an LLM as a general classifier for years. It's an incredibly common pattern for working with LLM judges or any sort of branched decision making workflow.

A lot of people who are a bit less familiar with the field saw the hype around Jev and presumed that the reason it was so exciting was that it was a fundamentally new interface for working with an LLM. And that additional excitement drove even more attention to Jev. But fundamentally, TypeSafe's announcement was that they found a particular architecture/training paradigm that resulted in a model for this particular interface that had incredible accuracy, very low latency, and for which they could offer inference at a super low cost.

I've not kept up with the flood of Jev clones that have been released, but I think this is just typical for any new component in deep learning that gets popular. There are an absurd number of open source autoregressive LLMs and fine tunes you can use. The thing that makes one more popular than the other is typically the general performance of the individual model.

But training a model for this purpose, or emulating the procedures described in Jev's papers, isn't something that would be beyond the capabilities of any lab. It's not an entirely alien architecture or approach.

The bigger question for TypeSafe as a company would be if other teams are producing Jev-like models that win on performance or cost. Like I said, I haven't followed the reports super closely, so no idea if that's the case or not.

Yes, it's easy. The thing people are missing (especially those believing AI is a "dead end" and "not transformative") is that the field has been advancing so fast in the past few years, that there's lots of such unexplored avenues, unpicked low-hanging fruits, that everyone just raced past. We've barely begun exploring the capabilities ML brought us - patterns, applications, and architectures.

Now that we're hitting against the hardware supply limits of global economy, I expect more people to go back and revisit the things left along the way in the mad rush to "just throw more compute at it / make a bigger model" - and thus many more cases like Jev to show up in the next few years.

Yes it's easy for an established shop, all they need to do is to tweak the post-training workflow. "Decision model" is the same kind of marketing as "LRM" attempted by OpenAI when RL CoT was new (to hyped up crowd). It's still fundamentally a classifier used for "decision making", games and RP were using generalist models and constrained outputs to do what the DOOM demo does for years.

The interesting part is also the easy part. The model and architecture are not hard for an experienced machine learning engineer to build.

The hard part is the data and evaluation. Sure, it’s not that hard to build a fast model with good predictive power. But fast at doing what? You probably don’t care about classifying whether a hotdog is a sandwich (which is the Jev demo).

Jumping on this, what about on-device?

I'd love an privacy first on-device model i could use in iOS.

We got you.

At0M: A 60M local Jev at 16 ms latency and 79% accuracy on Typed Decision

https://at0m.pienomial.com/ https://news.ycombinator.com/item?id=49920350

Well, for one, that one's not released as far as I can tell.

What’s the context window?

If it's 60M local, is it open source as well?

This account was created only a couple weeks ago and seems to be spamming this closed model, seems it might be a bot?

It hasn't been released yet, and local versions aren't always convenient

The concept existed a year before Jev or so. See Laya

You can use already trained large transformer models to make one, so it doesn't require the kind of high-scale compute, high quality data, data cleanup, reinforcement, and so on training that say an LLM does.

Smaller models have been able to do these sorts of tasks, but a little slower, for a while now. Give a small Qwen 3.8 model a classification task and force a structured output, and it'll do a good job. I've used Qwen 0.8b for basic image classification in <500ms on my local machine for a while now.

There are a few technical details that can reduce the latency significantly (covered in the post) but the real insight has been from watching the reaction to Jev and seeing that there's enough of a market interest to offer it as a distinct thing. The underlying concept/approach was already there.

Read this - https://magazine.sebastianraschka.com/p/classifier-history-a...

Transformers output a set of probabilities over outputs. For ChatGPT etc, those are predictions of what the next token will be. But it can also be a structured list of options or classes. Jev mostly innovated on the interface, API, and product concept around this, and made it click for a large number of people. Unfortunately for Jev, it's very easy to copy an API, and any pretrained LLM can be adapted to work in this way.

You just have to fine tune an LLM like Qwen on some synthetic data to do so. There was even someone that had a model that was exactly like Typesafe and published their work a year before Jev (but wasn't marketed as heavily since it was academic).

Jev created accessible/programmatic ergonomics around a general purpose classifiers, and did it very well; ie intuitive api and structured data approach.

Anyone can copy that and apply to an array of models - stripped down LLMs or already slim/highly performant traditional classification architectures (just wrap inference with an api that inputs/outputs the same structured data).

Jev, I think, would say their advantage is the intelligence of their models and training data including calibration: https://medium.com/code-applied/calibrated-classifiers-makin... (which i still struggle with in the general application... there's no free lunch with these things).

What's new is "smart" decision models than you can supposedly use on anything without additional training.

If you have a very narrow use case you can train a BERT based decision model on a laptop an hour if you have good data to train it on. It'll answer faster than the roundtrip to clef/jev and use <1gb memory

They are not too difficult to train if you already have infra to train regular LLMs. You can typically replace a few layers train them alone and you're off to the races.

Getting training data that works well for calibrated classification objectives is difficult.

I hear conflicting opinions (including my own) about how well calibrated each of these are. Jev seems to be the best.

But the jev release made obvious the PMF for these models, and the underlying reality is that calibration really doesn't matter much when you're replacing usecases where people were using damn LM head softmax probabilities before, which are nowhere near calibrated.

So now everyone simply finetunes qwen and makes a compared-to-regular-LLM vastly cheaper decision model. And it works for majority of usecases. People mostly only care about accuracy, not confidence.

The question of AI in automation is "can it make decisions in a consistent and predictable manner, with near 100% determinism?"

Many people seem to have run into the same question and started working out the answer.

It's not a new concept, it just took someone adding on to the approach and refining it. I never deep dove it, but I assume JEV is sort of like how Sora works? They had a blog post about how it has a sort of tiny LLM, which OpenAI's small LLMs are insanely good and well defined. I think any lab tackling this with a from-scratch model could yield affordable alternatives that are highly competitive.

It seems insanely obvious at least to me, that JEV is the new hot thing for the AI field since they give you stronger output that isn't... flat out wrong, that alone is impressive.

Most answers explain the LLM-based approach to these models, which is also what Typesafe did with Jev. However, depending on what you need, there are far simpler classification models, and for a lot of use cases, these models can be way faster and more accurate than Jev

But, for these adhoc models, you need to understand the task more, collect some data and train the model (on CPU, no need for GPU). So Jev-like models are a great way of getting a hosted general decision model, but if you have a very narrow task or set of tasks, you might be better off with some more basic models that you can run on the same server you run other things or even on your laptop