Been hoping for something in this space. Jev-like decision models on Qwen3.5 could really simplify some of our internal routing logic.

Yeah same. Got access to their API and then realised I don’t really have an immediate use case

You should call Jev-like models when you give it a JSON-like structure to produce, it is useful when you need _some_ intelligence in your code.

Edit: I want to add that you can see Jev like a smart if-statement.

To develop a smart ai system for my 2d roguelike platformer? game has way too many moving system for classic state-machine ai + i cant spare the time to develop it. its low latency entices me.

Consider every situation where you "force" an LLM to output only a choice / category, or a set of them. If you have workflows like that, you're now being promised significant cost- and latency reduction.

For coding agents it'd only be useful in a subset of situations. E.g. you could imagine using one to classify bash tool calls into safe and unsafe for example.

I don’t think this is a very good use case. You could do it better with a strong LLM and structured outputs.

The problem is that you want the model to carefully reason about the goal and code.

Zero shot classification with an approach like this isn’t going to do that. It’ll answer on first pass vibes.

Why not just let the LLM write a test instead of a "verifier"?

The git repo linked at top has some good examples for email classification for automatic email forwarding to specific departments, along with judging email tone and severity / priority, like for customer service emails.

Edit: The Flipper One is planning to have an LLM acceleration co-processor, and be able to host up to a 4GB VRAM size LLM. One use case they envision in their planning is using the microphone along with text to speech to be able to say, "Create an .ini file for this system with these specs" and the small LLM can do that on-device (its a handheld device) and then the user can use/send/upload that file.

Second Edit: I would love a mini LLM in KiCad or Altium that could take a component datasheet and produce a good footprint and schematic symbol for it.

Classifiers can be very useful for guiding the agentic loop which is basically a state machine. You have the agent propose a task, write some tests, write code, run tests, tests fail, write more code, go to acceptance, etc. So, a classifier can judge state transitions and decide what the agent should do next for example.

The part about "tool-calling on other models didn't solve already" is what gets you, sure I could tool call deepseek, glm or any other model, but the latency is huge and you get no confidence score. I gave JEV a shot via OpenRouter and it has a reply in less than 400ms, it's fast enough and cheap enough that you can hook it up to a game loop for example (so highly state dependant) and it can do decisions in real time.

It's appealing not having to fine-tune separate model for each use case

So you have more flexibility to get on with building, evolve your business logic etc

Arguably, it doesn't. Instead of a team of people doing break-fix on golden images you have that same size team of people doing break-fix on upstream playbooks. Lots of software doesn't achieve the goals but has a huge deployment story; not sure those things have ever been related.

Yeah except it doesn't really work

It does "work", you can download ansible today and use it, it does what it says. Is it the greatest solution for all use cases in infrastructure? Of course not, nothing is. Do people misuse it? Of course too, we're all human.

Regardless of what tooling you use, we're all building houses of cards, and depending on the situation, try to hold down those cards as well as we can, balancing a ton of other needs and requirements.

I would be curious to see comparisons of jev and similar things with problem specific classifiers. I think layla suggested making problem specific versions anyway? There is a lot of demand for magic don't do any work solutions, which is kind of weird in an era where agents can really help you build a customised solution effectively.

Agreed.

Just to be helpful if anyone is searching for layla, it's laya.

You already wrote the prompt, no? What I'd do, if I were you, is run the question through a LLM and then come back with targeted questions that it didn't answer.

I did the first part yesterday, jumped down the rabbit hole, and have 3 product ideas in my head now.

"Why is Jev fundamentally better than classification models like BERT or traditional ML?"

For a while is the keyword. It’s just vibe coders have just discovered the classifiers

Feels like astroturfing.

I think it’s timing. So many devs trying to squeeze their subscriptions, build more tooling to throughputMaxx. 6 months ago, i speculate it launches pretty flat.

Classifier models are extremely niche and trained for a singular purpose. A utility classifier that you can one-shot on almost any topic or need is a dramatically different beast.

Is it truly useful or accurate or beneficial? To be seen. But it's the idea that has everyone so captivated. An expert system that is an expert at most everything is a lot more useful than an expert system that is an expert at choosing a bar of soap, for instance.

Could you link to some of these classification models that can be used as versatile and perform with similar quality, speed and cost?

The Jev model is economically, but also in terms of compute, a much more efficient model. A normal LLM goes token by token, each token in a separate step. Whereas Jev just returns all the results the first round. So it is much better at classification than LLMs.

Compared to traditional ML classification, Jev works without training, like a LLM.

It’s simply not understanding. In terms of latency, LLM processing for large fact-based decision-making was incredibly slow. The real unlock here was speed in decision-making.

We know how useful classification models are, but massive pretraining is even more useful. So useful that it's tempting to just use a generic autoregressive model for everything, rather than trying to train a specialized classifier.

I think specialized models of all sorts, if it were possible to train them on equally much data, would outperform general models. In cases where we can have the data, like in self-play in games, they clearly do and have done so for a long time.

But it's a lot of effort to train specialized models. I think it's too early to tell if Jev delivers on its promises to give the best of both worlds, but if it doesn't, we'll have to make it work. It's just the right thing to try.

I guess you have AI to write your docker files and push your images now.

Are these forks? Different orgs doing the same thing as the OP?

If you have reusable components it should be quite easy for you to add it yourself. Its okay to code by hand, especially when doing frontend.

When comes to frontend work, the handmade UX designed by a professional will always outperform whats made by an llm.

Sure its 10x slower, but you'll get 20x more usere because it will obvious it was made with care (if done well, and you should learn to do it well). Its the same reason handmade clothing, watches and other luxery items speak to us more, sometimes its marketing but often times its the craft.

There's a reason why all the companies selling the idea that web dev is dead, like lovable, still hire designers with 50k retainers to do their landing pages.

I probably wouldn't be able to for data privacy reasons unfortunately, unless it's using an open-weight model (if it is, let me know). But I wish you the best of luck!

I noticed it has a pretty small context window of only 32k. For most tasks I guess it would be enough with ample context.

Are you on the foundation research team for Thomson? (If so, hiya from B!) Why would you expect Thomson to be particularly good at spam clf? I figured your additional corpus was all news and legal?

On Gemma 4 12B, I am getting 220 ms per move or QS. I used it to play the Snake game locally:

prompt_eval=244 ms wall=245 ms schema_cache=hit generated=0

Move limit reached after 200 moves: score=16, length=19.

So, if a 12B dense model can offer this latency on a local old PC, then definitely you can scale it up with more powerful machines and get even lower latency.

I can think of two possible approaches 1. Jev limits to 255 distinct options. So they can preprocess your set of options and “tell” the LLM via input tokens 1 = red, 2 = blue, etc then jev need only output softmax over 255 states while benefiting from pretrain of other LLMs 2. You allow the forward pass to output over the total token state but mask over the logits to limit to the user options. Less plausible? bc tricky when input is multi token which they clearly support.

My guess would be option 1. Didn’t read the kev repo here which would also explain

You can achieve open-vocabulary classification by making the final weights in the softmax come from a category encoder instead of being fixed learned weights. So instead of

softmax(encode(input)*learned_weights)

You have

softmax(encode(input)*encode(categories))

I'm not sure if Jev does it this way, but it's how you get open-vocabulary zero-shot image classification with models like CLIP [1].

[1] https://openai.com/index/clip/

Interesting approach with Qwen3.5 for decision models. Curious how "tiny" they've made them while keeping LLM reliability for critical paths.

Always?

The bright side of Jev being so popular could be that many companies and individuals realize that their applications might work well with a System One model, and they decide to run an open-source (or fine-tuned) version on their own

Why not name it Qev?

Oh Jared is cool - he made After and Razzle - nice

Also my question.

lol, this is hilarious

for some reason this is really funny to me. it's like the "black museum" black mirror episode where a consciousness in a toy animal can only communicate using very primitive predefined responses

Then why doesn’t jev identify as Claude?

Claude has identified itself as DeepSeek before if prompted in Chinese, is it DeepSeek?

No, it does not.

And Claude often identifies as Qwen or Deepseek when prompted in Chinese.

see sister comment's response https://news.ycombinator.com/item?id=49786548

Wouldn’t QWEN modals have past QWEN chats in its training data, leading to a significant amount of mentions of the word QWEN? Just the question “what model are you” would have been answered deterministically multiple times and they’re now part of the weights.

"<Q>What model are you?<A>Qwen." is surely in Qwen's training data. It's quite standard to include such meta knowledge during instruction tuning.

This is less true for modern posttrained models. Model identity can be explicitly reinforced during posttraining. Qwen's own finetuning docs include identity training examples, and Qwen models have been trained with system prompts that explicitly say things like "You are Qwen, created by Alibaba Cloud."

So a model correctly identifying its family doesn't necessarily mean it inferred that from pretraining.

I think with Jev, they took a posttrained model and trained it further, so it did not forget about its earlier knowledge during Owen's own RL.

This is, in fact, not true.

    $ ./build/bin/llama-cli -m Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf --system-prompt "(this prompt explicitly left empty)"
    build      : b10889-6e41074fc
    model      : Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf
    ftype      : Q4_K - Medium
    modalities : text
    using custom system prompt

    > What model are you?

    [Start thinking]

    Here's a thinking process:

    1.  **Analyze User Input:** The user asks "What model are you?"
    2.  **Identify Core Question:** The user wants to know my identity/model name.
    3.  **Check System Prompt/Instructions:** The system prompt says "(this prompt explicitly left empty)" - this is unusual but doesn't change my core identity. I need to respond based on my actual identity as Qwen / 通义千问.
    4.  **Formulate Response:** I should clearly state that I am Qwen (通义千问), developed by Alibaba Group's Tongyi Lab, and keep it concise as per guidelines.
    5.  **Draft Response:** I am Qwen (通义千问), a large language model independently developed by Alibaba Group's Tongyi Lab. How can I assist you today?
    6.  **Refine according to guidelines:** The draft matches the required identity and tone. It's direct and helpful. Ready.
    [End thinking]

    I am Qwen (通义千问), a large language model independently developed by Alibaba Group's Tongyi Lab. How can I assist you today?

Once the first letter is Q, the rest is probably pretty determined. Can you see the confidence for the first letter (don't want to accept the ToS to follow your link)?

If Jev is fundamentally trained using RLCD

Big if. More likely, it seems, is they started with an open LLM model and fine-tuned and repurposed it via their "RLCD" process.

We are indeed in an era of riches (thanks to the $1 Trillion a year being spent on this tech) that it is improving so fast I can just sit back with my 3 year old hardware and newer, better, more amazing workflows keep becoming possible for me just due to model compression / optimization + workflow developments. If I can't get something working this season, I just wait until 3 months from now and there will be an easier to set up, less resource demanding, better working version I can have my local AI install for me. Pretty wild times.

what kinda of specs would it need to run?

looks high lev

Distilling Jev should be super easy and cheap.

I tried to use jev for this. I'll share what I learned for the interested.

--

The setup was a simple map with different rooms. Each room had 1-3 doors.

For the inputs: The AI had an array of "known places" empty at the start, the current position, the current doors with no information about where they lead to, and the list of past actions

The goal / task for the AI was to explore all rooms and save them to known places. The AI needed to decide if to move or save the place at every turn.

--

So I wasn't able to make the AI explore all of the rooms. The AI kinda always wanted to move to the first option when moving. Out of 6 rooms it was able to save 3. My hypothesis is that jev as it is now is really bad at making connections and understanding it's input. So for example, even if it had a list of previous actions, it wasn't able to reason about it and know where to go. For this to work I'd need to explicitly tell it where it did not go. So you could say that the model is also not good with uncertainty / ambiguous scenarios.

edit: one last thing is that i replaced jev with an standard llm and it finished the goal no problem with the same information given

edit 2: it also felt like the same tradeoffs between small model vs large model. With small models you need to be very conscious and careful with the input while large models are more forgiving. Maybe jev is a small model, and we just need a larger one.

The open source ones- I downloaded a number and tried them and compared to Jev.

Anything that required knowledge / familiarity mmBERT and ModernBERT post-trains performed much worse.

So it seems like they did some kind of useful expansive pre-training.

Things that were Qwen or Gemma Diffusion did better at those kinds of tasks but were generally pretty inconsistent in terms of whether they could succeed repeatedly (and be stable + reliable) on the many types of tasks that are in the cookbook part of the Jev docs.

If you ask Jev similar input + questions, it's pretty stable. And does a reasonable job on a lot of questions.

This one public benchmark (the only I've seen) seems to give the open versions way too much credit. It wasn't my experience at all.

It gave my a false wrong sense of what might be required to get it working for something at work to avoid needing a new subprocessor as - at least on Cloudflare / OpenRouter Jev is third-party not hosted.

Shouldn't Jev-like models be built on top of diffusion models ? like GSAI-ML/iLLaDA-8B-Instruct ?

That showed us the best results at least

How is calibration of Jev or Jev-inspired models being evaluated?

You've got encoder and decoder reversed. BERTs and other models that primarily convert text (or other input) into latent representations are encoders. Models that convert their internal representations back into outputs are the decoders (in the case of GPT et al., autoregressive decoders because they perform this decoding based on past tokens).

You got it reversed. Bert is encoder only and gpt is decoder only.