The core idea is to craft the input prompt so that the first output token answers the question. This makes it possible to get a decision with a single forward pass.
In the blog post, we describe the approach in detail for GLM-5.3-Flash and vLLM. We benchmark this setup against Jev and Laya. We find that our setup is on-par with Jev in terms of accuracy and speed and that it substantially outperforms Laya.
Still, in terms of costs per decision, Jev is several x better than our setup. In turn, our setup supports vision inputs.
The Laya model compared in TFA shows one way Jev may be getting it's speed and low cost - by using a BERT-like bidirectional model rather than an auto-regressive one (LLM).
Off the top of my head, I would skip all the modern linear attention / state space stuff and use classical attention. But run prefill in a fully sliding-window mode so that “state” tokens simply don’t attend to far away tokens, or maybe also allow everything to attend to the first few tokens (and train like this). Now prefill is almost embarrassingly parallel, and you can make it fully parallel by duplicating work at block boundaries. (I’m not saying this is an awesome architecture if you want excellent results, but I’m also not convinced that Jev gives excellent results…)
The let queries attend to everything.
And architect the stack around this. Don’t try to cache the KV data — process the queries as you go so that the each input block and layer’s K and V data is computed, attended to, and discarded.
I’m curious whether Cerebras actually is a good device for this. Cerebras is kind of low on RAM, but if you don’t need to store KV data, maybe the entire computation fits on the die.
i have tested jev for my use cases and its horrendously wrong, but then the follow up from jev's team is "oh, you need to boil the question down further". it's a spiral of how much do you wanna dumb down the ask so that it answers it correctly. i'll pass for now.
also, 30k input tokens is a lot.
50k tok/sec is pretty impressive though.
But... when you were doing your measurements, were you using the same random book excerpt? If you were potentially getting even partial cache hits for your 50k tok/sec measurement, it would taint the benchmark: pretty much any inference provider running any LLM would be able to hit those numbers.
The prefill is the only blocking part, and you can prefill the whole context up to the point where they diverge, then prefill each question and decode the one token in parallel for each question.
If you batch vLLM calls with the same prompt prefix to the same process, it'll deduplicate the prompt prefix across batched requests (+/- the block size) and decode in parallel for each.
That's with a vanilla LLM. If you modify the LLM you can pull that in-graph, but it isn't really necessary.
Getting competitive accuracy with Jev is fairly easy, if by accuracy you mean that the highest weighted answer is the right one. GLM 5.3 is complete overkill, much smaller llms will do
What Jev brings to the table, beyond speed, is that the reported probabilities match actual likelyhoods. If you present three options, with A and B equally likely and C impossible, jev will approximately answer with 0.5, 0.5, 0. Stock LLMs don't
You can even do it with small not particularly hard to host local LLMs like a variant of Qwen 3.6 35B A3B or 3.8 27B.
https://developers.openai.com/api/docs/guides/structured-out...
Hallucinations in tool call results _do still exist_, but basically everyone asks for a JSON schema for the tool call and uses that to validate and re-prompt the LLM until it emits something with a valid schema.
That all goes out the window when a string field has a “hidden” schema in that only particular strings are valid, but that restriction isn’t in the JSON schema. I have had failures when I want a field to be specifically formatted Markdown or something.
We’ll probably see something that handles this better in the future like jsonnet or Cue or dhall that has some execution capabilities so you can write a custom validator beyond what JSON schema supports.
Well, that's essential what happens, but on the "backend" side still, so it's significantly more efficient. Also, model providers can play with e.g. how likely are they to accept a valid/invalid character, so for example the first token may only be [, { or ".
As for what's not in the schema, that's an orthogonal question.
Basically no one is doing that since 2024, have you been living under the rock? Read about constrained decoding.
> I have had failures when I want a field to be specifically formatted Markdown or something.
And Jev has an advantage here because it can’t generate anything?
I should also add that the source comes from one of about 400 possible places and in a variety of messed up formats, it's the raw feed from a news scraper...
Because it's proprietary? By using an open weight model you're guaranteed that you can access it forever; if one provider bans you then you can go to another one (or you can self-host). With a proprietary, single-provider model locked behind an API if your access is revoked you're screwed.
Fyi, I haven't tested this yet.
When it comes to the high volume market of business automation, it seems that ultra-low cost, rather than expensive frontier intelligence, is exactly what you want, and low latency is also nice to have for customer-facing applications like customer service chatbots.
When I hear “architecture” I am thinking number of parameters and latency.
When I hear “accuracy” I think training recipe, data, and (later) number of parameters.
So when you say that Jev’s architecture may not be necessary, the evidence I expect to see is comparable quality at comparable latency. Not equal quality at 2x latency and 4x the cost.
Hence my note about the risk of cost being the only moat.
But it is likely more than just a fine tune + novel training. At the very least the LM head is swapped out for a classifier one and then or also idk, bidirectional attention for the encoding pass I'm out of my depth at this point and will stop guessing. The training is probably where they have the biggest moat though, not that it's necessarily huge.
I have a project that fits jev as advertised almost comically well and I've been playing with it, and the various hacks and open versions. Jev doesn't necessarily perform better overall but it is quite different. It's sensitive to prompt phrasing in ways the others aren't, it's easy to generate questions where all the other models cluster in confidence but jev is an outlier. Not necessarily more correct, but it does feel like it's getting its answers in a different way.
I'm guessing just as much as anyone else but I've been spending a ton of time on this the last couple weeks, it landed right when I was most ready to dig into it.
I think I'm gradually getting a picture of what jev does, and why it might be good:
1) one shot classifiers are great, and are well-known.
2) LLMs are also great classifiers, but are slow at output - and not great at guaranteed structured output.
3) everyone focused on LLMs, and forgot about classifiers.
4) any decent one shot classifier should be easy to run "fast"/"in realtime".
5) combining 4) with LLMs and a harness (a loop) open up some clever use-cases that LLMs alone are too slow for.
As I understand it, jev isn't long for this world (we'll have foss one shot classifiers we can run on modest hardware) - but it seems likely the legacy will be that classifiers will re-gain some prominence alongside generative models.
What isn't obvious is why people keep shouting "Jev Jev Jev" all the time.
Astroturf.
I cannot fathom how people are not seeing the multiple daily posts as anything but the spam they are.
> As Privatemode is hosted in the EU and Jev is hosted in the US, we ran four of the datasets from Germany and from the US at the same time. From Germany, Privatemode answered in 180 ms and Jev in 264 ms. From the US, the order reverses: 164 ms for Jev against 299 ms for Privatemode.
turns out there is a trick to keeping the context filled and only evaluating a handful of choice tokens https://www.youtube.com/watch?v=bcGO7xre46o
I took a random book excerpt with 23,000 words (±30k input tokens) and used it as context. Jev still responds in 800ms, sometimes 500ms. That's in the neighbourhood of 20-50,000 tok/s prefill, which is obviously not possible with normal LLMs, not even Cerebras is this fast.