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.
If Jev is fundamentally trained using RLCD while you’re building on a Qwen model that was trained using RLHF, how can the resulting model be considered Jev-like?
Unless specifically told in a system prompt, the pile of weights has absolutely no knowledge of itself. You could hypothetically train it to answer such questions, but nobody bothers to do this, and ALL "knowledge" embedded in the weights is probabalistic anyway.
(I feel like this should be common knowledge in LLM discussions on HN by now.)
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.
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.
$ ./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)?
People seem to turn their brain off when it comes to this type of cargo culting. This doesn’t mean much. Qwen often identifies itself as Claude. Does that make it Claude?
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
TypeSafe is primarily a data research lab, which is how the biggest results in AI get made. We make all the data ourselves. We wouldn’t train on your data even if you asked us to (no offense). We do some pretty sophisticated stuff, but if you want to find out more, we’d have to hire you.
Bit of a Jev explosion going on. Is it because it's taking us back to a simpler time we understand better? Classification models have been around for a while.
>Classification models have been around for a while.
I'm still trying to catch-up on the Jev stuff, but my understanding is that it's basically just a more efficient LLM when all you want is the LLM to produce a classification.
There's more to it, of course, but it's not just "generic" classification ML because it accepts arbitrary inputs and can produce probabilities over arbitrary classes. Not saying this is the first time people have done this, but typically classification tasks are more static and limited.
In the same vein, it's also not just an LLM with structured outputs (which have been a thing for a while) specifically because that is a very inefficient way to approach classification using this kind of architecture. Jev models are much more performant because of how limited they are compared to a full LLM.
So when you want an LLM, but you only really need this kind of classification from the LLM, then Jev makes a ton of sense. This makes sense for me, since I've definitely used LLMs for this kind of classification work and, even then, it kind of felt like using a jackhammer to place some nails, etc.
Previous classification models need to be trained on the specific question/choices you are trying to output. Jev doesn't need to be retrained for every choice set provided.
LLMs can act as classifiers but they still have to generate text output in the form of a JSON object. This means they have to generate every single curly bracket, quote, command, etc. This turns out to be pretty expensive. On the other hand, Jev uses a different decision head so it doesn't generate text output at all, it outputs logits *only* for the choices provided. So it completely avoids the need to generate text at all, which means no malformed JSON and it's much faster as a result.
Finally, Jev also provides confidence scores that are actually reliable (not made up like LLMs).
The way I see this (I havent played around with Jev or layla the OSS version) is that classifiers have always existed and a recognised tool in the ML world. But, the norm is that one needs to not only know what to classify as, but determine what weights to use to classify the input.
Jev came in, and added that magic of "you dont need to train your classifier or determine the weights" if you dont want to, and just get the classified answer out. I think that's what is making people see this with a glitter in their eyes.
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.
It reminds me a bit of what Ansible got right: user communication. The underlying tech may have existed for a long time, but the genius is presenting it to a regular developer in a way that reads "yes, even you can understand ML, just using a little JSON". The contribution of that should not be understated, as has been clearly evident recently.
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.
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.
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.
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.
> and enabled things that were impractical previously
I think that there are not _that_ many use-cases that have been opened up by this that tool-calling on other models didn't solve already. Really depends what benchmark you're looking at. This one against BANKING77[0] has many issues, but suggests it's really not far off DeepSeek 4.1 Flash. This one against BoolQ[1] shows marginal improvement over Qwen3.6. This one against MMLU-Pro[2] (same author as the previous) shows significant improvements over two Qwen models.
So there's definitely _some_ alpha there, but I don't think it's the sea-change that the hype would suggest; that is to say, yes, some things that weren't practical before are now, but many things were already very practical with the existing tools.
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 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.
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.
Jev is creating a sort of identity crisis for me, because the number of absolutely clueless folks parroting the classifier thing is the first time I've seen this sort of mass psychosis in CS upfront.
Like even 5 minutes of tinkering captures why this isn't anymore like BERT or any past classification model than ChatGPT is like those old Markov Chain generators, yet folks cannot shut up about how this is nothing new.
Absolutely scary and makes me wonder how much of the field is just people super confidently discrediting otherwise promising/interesting directions for development for a cheap dunk!
Show me a single example how is this jev thing better than a modern Bert solution?
Or even llm if you claim about versatility. You can easily modify the llm inference code to make it predict a single token represent the classification choice and extract the probability that way.
Sure jev will still be faster, but a local deployed Bert model is way faster than both.
And to get the most out of it you still need to fine tune the models anyway, unless your classification task is just one of those mainstream ones.
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?"
I think a great use case for these will be when they have large context windows and are able to enforce styling rules for frontend development, and component creation rules for react. You can then ditch the styles guides and styling skills and create a decision tree for enforcing styling, so that you can't run into drift issues or duplication issues. That's where I'm wasting most of my time right now, constantly correcting all of the UX/UI issues that are created for every single feature.
Back in the good old days we would prevent these ux/ui issues by rigorously enforcing the use of our own stylesheets and classes. Later that grew to only using the company ux components. This was very successful in keeping everything neat and tidy. The only drawback was creating and curating new elements and getting consensus. But otherwise it works wonders.
Try constructing reusable components out of what you are doing instead of building everything up from basic building blocks. This also allows more concrete testing of individual parts and then if you want to change the look you can change it in one place and have it apply everywhere.
Agentic development doesn’t mean “throw all what we learned out of the window”, the same practices that helped speed up and improve quality of work of humans also helps agents. In fact, the multiplier is even bigger. You will notice it in development speed and reduced cost due to avoiding churn.
I use reusable components but the issue is that the model rarely checks to see if a component already exists. Or it will use the wrong one. Or if the component strays from the reusable component it doesn't extend it in a generalized way and will special case inline styles, or it'll create a new component. Rarely does it intelligently figure out the correct course of action.
I've also built up a suite of linter rules to catch the same mistakes the model makes over and over. Still, there are a lot of gaps. I think it's mostly because my codebase is massive at this point. It was easy when the codebase was small and didn't require context gathering to make good decisions.
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.
Hi, I'm currently planning on launching a product like this "Grammarly for Design" in a few weeks. Would you be interested in being part of the alpha group ?
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 hope these get small and good enough to create “pet like” AIs for games. You know, like scream “follow me” at an NPC, STT stack translates it and feeds it to a local Jev-like model that then picks a number of things for the NPC to do.
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.
Is Jev a decoder (e.g., BERT) or is it some kind of encoder (e.g., GPT) that just happens to be trimmed down to only outputting a handful of tokens for the answers and their probability?
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).
Quite impressed by the energy people are putting into making OSS Jev-like models.
I understand the hype but I wonder: what are the use cases for this kind of model? Could it be used in the context of coding agents, or is it more relevant in totally different situations?
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.
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.
> what are the use cases for this kind of model? Could it be used in the context of coding agents
Yeah, it could. The most obvious usage would be to have local fast cheap "feedback" / "control" over a slower more expensive agent (i.e. cc / codex / opencode). Things like "goals" could now be split from a long prompt into "actions" and "verifiers". Where for each action you also produce a verifier. Then after each action you run the verifier w/ this kind of "universal classifier" and decide if the step was done correctly, if it needs follow-up and so on.
Example: implement auth in this repo -> llm_plan() -> for item in plan generate_verifier() -> for item in plan implement() ; verify() ; accept() / followup().
Verifiers could be something like this. take a plan item as input, generate classification questions that might verify the task "is this following project conventions?" | "is this touching files from other tasks?", etc.
You can do that with LLMs, but some things might become cheaper / faster. And you can pretty much use it to check against an ever growing list of conventions. Yours or project specific.
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.
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.
I wonder how these would do filtering my spam. I have been using 27B-class models for a while now, and they are nearly perfect at determining what is spam and what isn't. The only disadvantage is computational cost.
Take a look at Thomson 1.0-small, which is a variant of qwen 3.6 35b post trained by Thomson Reuters for text analysis. It classifies text content very well.
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.
Because these decision models do not have tool calling, the knowledge cutoff might become a problem. We'll either have to keep training continuously if we run locally or switch to the newer version every month or so when using a closed one like Jev
even with knowledge cutoff set a second from now, you still want to provide as much info as you can if you’re using such tools for delegating decisions
Can someone do a ELI5A of how they achieve classification over any user defined list of items? Normal neural networks do a softmax over a known output set to get probabilities
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].
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
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
All the people that are just writing an Jev-like API on top of a normal LLM are missing the point. What makes Jev special is the training data; it's how it's trained. The architecture is probably nothing special. Just a text encoder with parallel prediction branches.
I have tried many of these open-source Jev-like models on some linguistic tasks and they are so bad compared to Jev.
I hope so. And I would really like to try an actual Jev open source model. But it will make it more difficult to market it when someone releases something like that because of so many of these "open source Jev-like model".
I'm just sitting back for a few weeks / a couple months to let it shake out, let others put in all the work, and then see if people are still interested and finding use cases that this access model fits better than the usual chat completions endpoint people are used to.
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.
Found this benchmark for Jev-class models: https://benchmarkheaven.com/jev-models
There are already many Jev-like models in there.
Edit: No affiliation. Just found it and thought others might find it interesting.
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.
If Jev is fundamentally trained using RLCD while you’re building on a Qwen model that was trained using RLHF, how can the resulting model be considered Jev-like?
I can't find it but saw that if you give Jev English alphabet as choices and ask it in a loop what model it is, it would say Qwen
also tried myself: https://console.typesafe.ai/playground?share=shr_1690a3160f1...
That is not how models work.
Unless specifically told in a system prompt, the pile of weights has absolutely no knowledge of itself. You could hypothetically train it to answer such questions, but nobody bothers to do this, and ALL "knowledge" embedded in the weights is probabalistic anyway.
(I feel like this should be common knowledge in LLM discussions on HN by now.)
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.
"<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.
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.
This is, in fact, not true.
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)?
People seem to turn their brain off when it comes to this type of cargo culting. This doesn’t mean much. Qwen often identifies itself as Claude. Does that make it Claude?
And Claude often identifies as Qwen or Deepseek when prompted in Chinese.
see sister comment's response https://news.ycombinator.com/item?id=49786548
Sort of implies its basis, doesn’t it?
Claude has identified itself as DeepSeek before if prompted in Chinese, is it DeepSeek?
Then why doesn’t jev identify as Claude?
No, it does not.
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
lol, this is hilarious
> 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.
The FAQ on the Jet announcement (https://typesafe.ai/blog/introducing-system-one-models-and-j...) claims it was trained with their own data:
Where Does Our Training Data Come From?
TypeSafe is primarily a data research lab, which is how the biggest results in AI get made. We make all the data ourselves. We wouldn’t train on your data even if you asked us to (no offense). We do some pretty sophisticated stuff, but if you want to find out more, we’d have to hire you.
Also my question.
Bit of a Jev explosion going on. Is it because it's taking us back to a simpler time we understand better? Classification models have been around for a while.
>Classification models have been around for a while.
I'm still trying to catch-up on the Jev stuff, but my understanding is that it's basically just a more efficient LLM when all you want is the LLM to produce a classification.
There's more to it, of course, but it's not just "generic" classification ML because it accepts arbitrary inputs and can produce probabilities over arbitrary classes. Not saying this is the first time people have done this, but typically classification tasks are more static and limited.
In the same vein, it's also not just an LLM with structured outputs (which have been a thing for a while) specifically because that is a very inefficient way to approach classification using this kind of architecture. Jev models are much more performant because of how limited they are compared to a full LLM.
So when you want an LLM, but you only really need this kind of classification from the LLM, then Jev makes a ton of sense. This makes sense for me, since I've definitely used LLMs for this kind of classification work and, even then, it kind of felt like using a jackhammer to place some nails, etc.
Happy to be correct, though.
Previous classification models need to be trained on the specific question/choices you are trying to output. Jev doesn't need to be retrained for every choice set provided.
LLMs can act as classifiers but they still have to generate text output in the form of a JSON object. This means they have to generate every single curly bracket, quote, command, etc. This turns out to be pretty expensive. On the other hand, Jev uses a different decision head so it doesn't generate text output at all, it outputs logits *only* for the choices provided. So it completely avoids the need to generate text at all, which means no malformed JSON and it's much faster as a result.
Finally, Jev also provides confidence scores that are actually reliable (not made up like LLMs).
The way I see this (I havent played around with Jev or layla the OSS version) is that classifiers have always existed and a recognised tool in the ML world. But, the norm is that one needs to not only know what to classify as, but determine what weights to use to classify the input.
Jev came in, and added that magic of "you dont need to train your classifier or determine the weights" if you dont want to, and just get the classified answer out. I think that's what is making people see this with a glitter in their eyes.
Agreed.
Just to be helpful if anyone is searching for layla, it's laya.
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.
It reminds me a bit of what Ansible got right: user communication. The underlying tech may have existed for a long time, but the genius is presenting it to a regular developer in a way that reads "yes, even you can understand ML, just using a little JSON". The contribution of that should not be understated, as has been clearly evident recently.
Yeah except it doesn't really work. It constantly breaks underneath you. The whole system has to be managed, e.g NixOS, or else it's a house of cards.
I'm also not personally a fan of Ansible, but to claim it doesn't really work is quite breathtaking given the size of the installed base.
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.
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.
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.
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
For a while is the keyword. It’s just vibe coders have just discovered the classifiers
It’s because it’s practically useful and enabled things that were impractical previously.
> and enabled things that were impractical previously
I think that there are not _that_ many use-cases that have been opened up by this that tool-calling on other models didn't solve already. Really depends what benchmark you're looking at. This one against BANKING77[0] has many issues, but suggests it's really not far off DeepSeek 4.1 Flash. This one against BoolQ[1] shows marginal improvement over Qwen3.6. This one against MMLU-Pro[2] (same author as the previous) shows significant improvements over two Qwen models.
So there's definitely _some_ alpha there, but I don't think it's the sea-change that the hype would suggest; that is to say, yes, some things that weren't practical before are now, but many things were already very practical with the existing tools.
0: https://sanand0.github.io/llmevals/jev/
1: https://github.com/ekzhang/openjev-sglang/blob/a3554ed9e9c26...
2: https://github.com/ekzhang/openjev-sglang/blob/a3554ed9e9c26...
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 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.
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.
Could you link to some of these classification models that can be used as versatile and perform with similar quality, speed and cost?
Feels like astroturfing.
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.
Jev is creating a sort of identity crisis for me, because the number of absolutely clueless folks parroting the classifier thing is the first time I've seen this sort of mass psychosis in CS upfront.
Like even 5 minutes of tinkering captures why this isn't anymore like BERT or any past classification model than ChatGPT is like those old Markov Chain generators, yet folks cannot shut up about how this is nothing new.
Absolutely scary and makes me wonder how much of the field is just people super confidently discrediting otherwise promising/interesting directions for development for a cheap dunk!
Show me a single example how is this jev thing better than a modern Bert solution?
Or even llm if you claim about versatility. You can easily modify the llm inference code to make it predict a single token represent the classification choice and extract the probability that way.
Sure jev will still be faster, but a local deployed Bert model is way faster than both.
And to get the most out of it you still need to fine tune the models anyway, unless your classification task is just one of those mainstream ones.
Hey I am clueless, how do I learn more?
Why is Jev fundamentally better than classification models like BERT or traditional ML?
Happy to read a written response or if you suggest a prompt to put into my LLM to get it to research and explain the relevant details.
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?"
I think a great use case for these will be when they have large context windows and are able to enforce styling rules for frontend development, and component creation rules for react. You can then ditch the styles guides and styling skills and create a decision tree for enforcing styling, so that you can't run into drift issues or duplication issues. That's where I'm wasting most of my time right now, constantly correcting all of the UX/UI issues that are created for every single feature.
Back in the good old days we would prevent these ux/ui issues by rigorously enforcing the use of our own stylesheets and classes. Later that grew to only using the company ux components. This was very successful in keeping everything neat and tidy. The only drawback was creating and curating new elements and getting consensus. But otherwise it works wonders.
Try constructing reusable components out of what you are doing instead of building everything up from basic building blocks. This also allows more concrete testing of individual parts and then if you want to change the look you can change it in one place and have it apply everywhere.
Agentic development doesn’t mean “throw all what we learned out of the window”, the same practices that helped speed up and improve quality of work of humans also helps agents. In fact, the multiplier is even bigger. You will notice it in development speed and reduced cost due to avoiding churn.
I use reusable components but the issue is that the model rarely checks to see if a component already exists. Or it will use the wrong one. Or if the component strays from the reusable component it doesn't extend it in a generalized way and will special case inline styles, or it'll create a new component. Rarely does it intelligently figure out the correct course of action.
I've also built up a suite of linter rules to catch the same mistakes the model makes over and over. Still, there are a lot of gaps. I think it's mostly because my codebase is massive at this point. It was easy when the codebase was small and didn't require context gathering to make good decisions.
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.
Hi, I'm currently planning on launching a product like this "Grammarly for Design" in a few weeks. Would you be interested in being part of the alpha group ?
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 hope these get small and good enough to create “pet like” AIs for games. You know, like scream “follow me” at an NPC, STT stack translates it and feeds it to a local Jev-like model that then picks a number of things for the NPC to do.
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.
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
Half-OT:
Is Jev a decoder (e.g., BERT) or is it some kind of encoder (e.g., GPT) that just happens to be trimmed down to only outputting a handful of tokens for the answers and their probability?
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.
Quite impressed by the energy people are putting into making OSS Jev-like models.
I understand the hype but I wonder: what are the use cases for this kind of model? Could it be used in the context of coding agents, or is it more relevant in totally different situations?
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.
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.
> what are the use cases for this kind of model? Could it be used in the context of coding agents
Yeah, it could. The most obvious usage would be to have local fast cheap "feedback" / "control" over a slower more expensive agent (i.e. cc / codex / opencode). Things like "goals" could now be split from a long prompt into "actions" and "verifiers". Where for each action you also produce a verifier. Then after each action you run the verifier w/ this kind of "universal classifier" and decide if the step was done correctly, if it needs follow-up and so on.
Example: implement auth in this repo -> llm_plan() -> for item in plan generate_verifier() -> for item in plan implement() ; verify() ; accept() / followup().
Verifiers could be something like this. take a plan item as input, generate classification questions that might verify the task "is this following project conventions?" | "is this touching files from other tasks?", etc.
You can do that with LLMs, but some things might become cheaper / faster. And you can pretty much use it to check against an ever growing list of conventions. Yours or project specific.
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"?
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.
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.
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.
Yeah same. Got access to their API and then realised I don’t really have an immediate use case
I wonder how these would do filtering my spam. I have been using 27B-class models for a while now, and they are nearly perfect at determining what is spam and what isn't. The only disadvantage is computational cost.
Take a look at Thomson 1.0-small, which is a variant of qwen 3.6 35b post trained by Thomson Reuters for text analysis. It classifies text content very well.
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.
How is calibration of Jev or Jev-inspired models being evaluated?
Because these decision models do not have tool calling, the knowledge cutoff might become a problem. We'll either have to keep training continuously if we run locally or switch to the newer version every month or so when using a closed one like Jev
even with knowledge cutoff set a second from now, you still want to provide as much info as you can if you’re using such tools for delegating decisions
I noticed it has a pretty small context window of only 32k. For most tasks I guess it would be enough with ample context.
Can someone tell me what is the difference between Jev and a normal neural network that does classification ?
My understanding is: it takes text input and it does one shot classification (no training data)
Yes, this is essentially it.
As a corollary, the output classes can be any set, rather than needing to be set before training.
Can someone do a ELI5A of how they achieve classification over any user defined list of items? Normal neural networks do a softmax over a known output set to get probabilities
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/
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
Been hoping for something in this space. Jev-like decision models on Qwen3.5 could really simplify some of our internal routing logic.
- https://github.com/logan-markewich/jeff - https://github.com/bespokelabsai/nimble
Are these forks? Different orgs doing the same thing as the OP?
Why does nobody ever ship these as a docker image?
I guess you have AI to write your docker files and push your images now.
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
All the people that are just writing an Jev-like API on top of a normal LLM are missing the point. What makes Jev special is the training data; it's how it's trained. The architecture is probably nothing special. Just a text encoder with parallel prediction branches.
I have tried many of these open-source Jev-like models on some linguistic tasks and they are so bad compared to Jev.
It won't be long until people produce a decent training data set generation pipeline.
The number of people working on this is crazy. Something will coalesce.
I hope so. And I would really like to try an actual Jev open source model. But it will make it more difficult to market it when someone releases something like that because of so many of these "open source Jev-like model".
I'm just sitting back for a few weeks / a couple months to let it shake out, let others put in all the work, and then see if people are still interested and finding use cases that this access model fits better than the usual chat completions endpoint people are used to.
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?
Why not name it Qev?
Interesting approach with Qwen3.5 for decision models. Curious how "tiny" they've made them while keeping LLM reliability for critical paths.
Distilling Jev should be super easy and cheap.
Interesting to see a Jev-like approach applied to Qwen3.5. Always appreciated Jev's simplicity for quick decisions.
Always?
looks high lev
Oh Jared is cool - he made After and Razzle - nice