TypeSafe AI's Jev sparks priority fight with open-source rival Laya
A week after TypeSafe AI unveiled its fast, text-free 'System One' model Jev, an independent developer says he published the same architecture a year earlier and released a rival open model, Laya.
What to know
- Jev is a new model class that outputs typed, calibrated probabilities instead of text, claimed to be 20-200x faster and 40-400x cheaper than comparable LLMs.
- Enterprise adoption was unusually fast per Vercel's AI Gateway data, but early access required a waitlist and lacked technical papers or docs.
- An open-source developer says he published the same decision-model architecture a year earlier with open weights and papers, and released a rival model, Laya, after TypeSafe launched without crediting him.
- Independent benchmarks, reverse-engineered clones and browser-based comparison tools (JevBench, OpenJev) are now testing whether Jev's speed and accuracy claims hold up.
The dispute Whether TypeSafe's Jev represents a genuine breakthrough or an uncredited repackaging of the open Laya author's year-old research is actively disputed. · positions read across 205 posts and comments
Jev's decision-model design is a genuinely useful new shape for classification and agent workflows.
-
“jev is basically a general purpose classifier. and i think that is amazing. super curious what the training set for that looks like.”
@badlogicgames · X ↗
The launch is overhyped relative to how closed and undocumented it actually is.
-
“Lots of hype on this lately but I am honestly disappointed that I have to join a wait-list and still can't even see the doc to try to understand what the capabilities are and how are we supposed to use them...”
loige · Lobsters ↗
TypeSafe should have credited prior open-source work describing the same architecture.
-
“I think their problem is more not being cited by the team at typesafe, as in general academic politeness.”
porridgeraisin · Hacker News ↗
Diogo Almeida Founder, TypeSafe AInandakishor_ml Creator of Laya (open-source rival)Simon Willison Independent AI commentatorflorianstandhar Creator of JevBench
How it unfolded 6 developments, newest first · click a bar or a number to jump articlespostscomments
-
6
TypeSafe founder reframes Jev as a production tool, not a breakthrough oracle
In a video interview, TypeSafe's creator pushed back on runaway hype around Jev, describing System One models as built for production automation rather than as a general intelligence leap.
“I'm with Maggie Appleton, I think "decision models" is a better name for these…”
— Simon Willison, independent AI commentator · source -
Can we all agree to ban the term 'non-autoregressive model' from our collective vocabulary? It's like naming a database a 'non-spreadsheet'. Decision models, direct inferenceing models, etc... there are sooo many better names for this. I would have thought we had learned our lesson with the whole 'Vectorless RAG' incident. We should go back and…
2 more of the top 3 · 5 posts in this stretch
-
This is the one claims not to be a llm I think, that's all I know about it however.
-
Looks cool! Is there any relation to this area: "Type-constrained code generation with language models"?
-
- 2 days quiet
-
5
Open-source rival Laya launches; creator accuses TypeSafe of uncredited copying
Developer nandakishor_ml released Laya, an open-weight 'System 1' decision model he says runs 6-8x faster than Jev, while publicly stating that he had published the same non-autoregressive decision-model concept, with papers, open weights and datasets, a year before TypeSafe's launch and was never credited.
“They proposed the exact same non-autoregressive decision concept as if it was a brand-new scientific breakthrough. Except they launched without technical papers, without open weights, and with zero open training datasets.”
— nandakishor_ml -
N
🚀 Next Tool Tech Roundup · Sep 20 Jev (TypeSafe AI) is a non-LLM transformer that outputs probabilities, not text. Built by an OpenAI RLHF co-inventor. Can't hallucinate. Vercel reports 5–18× faster than OpenAI for classification.
2 more of the top 3 · 64 posts in this stretch
-
The unfortunate true is that getting even the best work in front of an audience is often much harder than solving the problem. Is uploading a paper to arXiv enough to expect the work to be recognized and cited? Unfortunately, it rather is not. arXiv is an open repository which includes plenty of not reviewed and not officially published papers. In…
-
@
Jev is the AI model for the token-pinching era. — Instead of generating text, it makes decisions: yes/no, where to route this request, whether to escalate this issue, etc. — The key idea is that not every AI task needs an LLM generating text. Nearly 13% of Vercel's paid AI teams tried it in 24 hours. …
-
-
4
Vercel reports record-fast enterprise adoption of Jev
Vercel said Jev was adopted faster than any other model in its AI Gateway's history, reaching roughly 13% of teams on day one, about 2x the rate of the GPT-5.6 family and 6x that of Fable 5.1.
“Jev was adopted faster than any other model in AI Gateway history.”
— @vercel -
first by Vercel, 3d ago
-
Jev was adopted faster than any other model in AI Gateway history. In the first day, @typesafeai reached ~13% of teams, 2x the GPT-5.6 family and 6x Fable 5.1.
2 more of the top 3 · 9 posts in this stretch
-
Yeah, that’s legit impressive and immersive
-
DebertaV3's architecture and noising should be even better as a basis because it had a couple inductive biases (cross encoder, disentangled attention and RTD corruptions) that enabled it to have unmatched weight performance ratio on such tasks.My gut tells me that a better approach to a calibrated 0-shot classifier than shoehorning DiffusionGemma…
-
-
3
Independent developers launch OpenJev, a browser-based comparison tool
OpenJev let users run small open local models (e.g. MiniCPM5, Qwen3) in-browser and compare their direct-probability and generated-JSON outputs against Jev's published accuracy figures, aiming to make Jev's claims independently testable.
“jev is basically a general purpose classifier. and i think that is amazing. super curious what the training set for that looks like.”
— @badlogicgames, X user · source -
jev is basically a general purpose classifier. and i think that is amazing. super curious what the training set for that looks like.
2 more of the top 3 · 40 posts in this stretch
-
The (lack of) accessibility of this website is beyond words.
-
You might also be interested in "Open-sourced jev architecture last year with model,paper and dataset"
-
- 1 day quiet
-
2
AI developers react, calling Jev a powerful classifier and starting reverse-engineering
Commentators on X and HN reacted with a mix of excitement about the classifier design and skepticism about the extraordinary claims, while some began reverse-engineering a 'jev-like' architecture and posting training code.
-
Sounds almost too good to be true. Jev is an AI model built for decisions rather than text generation, introduced by Diogo Almeida and trained using a new method called RLCD. Claimed to be 20–200× faster and 40–400× cheaper than comparable LLMs. Costs $0.042 per million input
2 more of the top 3 · 37 posts in this stretch
-
T
TypeSafe AI debuts Jev, a model using "Reinforcement Learning for Calibrated Decisions" to produce typed probabilistic decisions that software can use directly (Thomas Claburn/The Register) https://www. theregister.com/ai-and-ml/2026 /09/16/typesafe-ai-debuts-model-for-machines-that-plays-doom/5296711 http://www. techmeme.com/260916/p1#a260916 p1
-
> Structured outputs slot into ordinary software as fuzzy decision rules: classify, route, score, extract, or branch where hand-written logic is too brittle.Oh, I have one of those use cases, matching people in genealogy trees. You can ask all sorts of questions: do the names match? Do they match within some edit distance? Do they match according…
-
-
1
TypeSafe AI launches Jev, a text-free 'System One' model
TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, released Jev in early access: a model that takes text input but outputs typed, calibrated probabilities instead of strings, trained with a new method called Reinforcement Learning for Calibrated Decisions (RLCD), claimed to be 20-200x faster and 40-400x cheaper than comparable LLMs at $0.042 per million input tokens.
“I am beyond excited to announce that today, TypeSafe AI is releasing our first System One Model…”
— Diogo Almeida -
first by AIModels.fyi, 3d ago · also Arize AI, MarkTechPost
2 more headlines
- TypeSafe's Jev: Can decision models replace LLM judges? Arize AI · 3d ago
- TypeSafe AI Releases Jev: A System One Model That Returns Typed, Calibrated Decisions Instead of Text MarkTechPost · 3d ago
-
first by Simon Willison, 2d ago · also HN Frontpage
1 more headline
- Jev introduces a new shape of LLM HN Frontpage · 2d ago
-
D
"After two years in stealth, countless technical challenges, and research breakthroughs… I am beyond excited to announce that today, TypeSafe AI is releasing our first System One Model: a new class of frontier models built to make fast, structured decisions that software can use directly." https:// typesafe.ai/blog/introducing-s…
2 more of the top 3 · 50 posts in this stretch
-
This, combined with contracts, could make a lot of things so much fun now!For those who don't know (which is probably everyone but me), I ported the design-by-contract pattern in Python and combined it with LLMs. This was early 2025. I originally wrote about it here: https://leoveanu.com/2025-03-01-dbc/ . Contracts are a core feature of SymbolicAI…
-
Another brilliant launch for developers: and its 20-200x faster than LLMs because it skips token-by-token generation entirely. TypeSafe AI just launched Jev, > 20-200x faster >40-400x cheaper (w/ output tokens free) > Frontier composable intelligence optimized for decisions So Jev is an AI model...
-
What people are saying 10 voices from 4 sites · best of 205 · verbatim
- What does Jev's actual training set and data pipeline look like?
- Will TypeSafe release technical papers, open weights, or training data to substantiate its claims?
- Sep 20
-
Jev is not open-source and only available via API. Here's an open version called Nimble which performs just as well.
-
H
Some jev-like alternative server benchmarks (many run local)
- Sep 19
-
I think their problem is more not being cited by the team at typesafe, as in general academic politeness. On the one hand you have the charitable assumption that they developed it independently. On the other hand, my opinion is that it is naive to expect companies to do that even if they took inspo from it, especially when this is a core product…
-
OP's was leaky slop from day one [0][1], as is his article [2]It is arrogant and entitled for the author to take credit for the concept of RL over sequence embeddings, and none of the work that went into pretraining, not to mention the egregious target leakage [1][0]: Author fails to grasp the concept of virtual environments…
-
I think the biggest lesson with Jev was the one of communication and understanding for the broader audience, sometimes a lot about innovating involves repeating yourself and translating your own thoughts to an intended audience.Classical machine learning has been, for the most part, and just by the nature of science, behind academic terms and…
- Sep 18
-
Thanks a million for both the insights! ... I could have probably guessed that URL in retrospect :'D
-
> see the doc https://docs.typesafe.ai/ are the docs. You can't find it from the website... > have to join a wait-list The model is available on Vercel and Openrouter without waitlist. (Waitlist wait times appear to be between a few minutes and a day.)
-
Lots of hype on this lately but I am honestly disappointed that I have to join a wait-list and still can't even see the doc to try to understand what the capabilities are and how are we supposed to use them...
- Sep 17
-
- Jev is a very good, cheap, fast classifier - incorporating it lets you use more workflow, DAG, control flow primitives in your agent flows …. LangGraph
- Sep 16
-
I reverse-engineered a jev-like architecture given its type. You can find the repo here to train your own jevlikes: