Jev for research and data
Benchmarks people ran, architectures people reverse-engineered, results that did not work.
20 items, from every source we cover All use cases
-
Hacker News · thread
Benchmarking Jev: what a decision model can (and can't) do in an agent harness
The most useful dev.to piece on this model so far: about 22,500 calls across ten public datasets, with the code and raw results in a companion repository. Half the value is the negative results — model-difficulty routing at 51% (no signal at all), trajectory failure attribution at 0.560 AUROC (random), and non-English retrieval well below English. The other half is two engineering lessons from iterating: decompose a fuzzy judgment into orthogonal questions and compose them in code, and let a choice compete before a noul verifies.
- X
X · image
Build a Jev Judge
Takes the same argument to evaluation: if the judgement is a handful of bounded decisions, does it need another round of text generation to make it? The most concrete Jev-as-judge case we have found on X, and the one we would test before relying on it.
- GitHub
GitHub · project
featherless-ai/simple-jev
A different route to the same output contract: read the next-token logits for each question and assemble choices, rubric scores or truth values, with no classifier head and no training. Served as a playground, which makes it the quickest way to see the shape of a Jev answer without a key.
- GitHub
GitHub · project
jaredpalmer/kev
The Kev family: the same decide-and-act shape rebuilt on Qwen3.5 instead of the hosted API. It is also the build whose Hacker News thread trended, and the clearest sign that the pattern travels between models.
- YouTube
YouTube · video
Jev Explained: Demos and Use Cases
Demo-heavy and grounded in concrete uses rather than architecture, which makes it the better second watch after the walkthrough above.
- X
X · image
Jev might genuinely be an “Internet moment” for AI. TypeSafe reports up to 193x faster and 44…
Quotes the vendor's up-to-193× faster claim while framing it as the maker's number. We have not reproduced that measurement, and the row says so rather than repeating it flatly.
-
X · post
Jev 发布没几天，开源社区已经开始疯狂复刻了🔥 最值得推荐的五个模型： 1、Laya 421M：原生决策模型，支持 Mac 2、Decider-2B：最像 Jev，基于 Qwen3.…
A Chinese-language roundup of five open replications, with parameter counts and platforms — evidence for the copy-the-pattern claim that the Kev repository makes alone.
-
Hacker News · thread
Jev, Prolog, Pi, and the dream of probabilistic logic programming
No comments yet, which is itself the interesting part: an essay putting Jev next to Prolog and probabilistic logic programming — the argument that a decision is a query, not a completion. Read it for where the idea came from.
-
X · post
Jevの内部アーキテクチャを推測している技術記事(Jev’s Architecture Unmasked)からメモ。 ・本記事はJevのAPIを約1万回の呼び出して、内部構造を推測したも…
A Japanese engineer's notes on an article that reverse-engineered the architecture from about ten thousand API calls. The closest thing here to the architecture conversation.
- Hacker News
Hacker News · thread
Kev: Tiny Jev-like family of decision models built on top of Qwen3.5
A third-party rebuild of the same idea on Qwen3.5 instead of a call to the hosted model. The clearest evidence that what people are copying is the decide-act loop, not the endpoint.
- GitHub
GitHub · project
kitfunso/hippo-memory
The most carefully reported Jev result in this list, including its own negative: an opt-in reranker takes R@1 from 0.41 to 0.62 on a private 300-query store at about $0.0004 a recall, the margin held in 20 of 20 seeds — and three graded tests still did not beat the free local cross-encoder. Off by default.
- YouTube
YouTube · video
Open Jev Models Are Here!!
Runs seven open Jev-style models against a public benchmark rather than explaining the API again, which makes it the most useful of the second wave. Sponsored by Dell for the compute and the sponsorship is stated in the description; the models named there are the ones we list in the builds column.
- Hacker News
Hacker News · thread
Open-sourced jev architecture last year with model,paper and dataset
A prior publication arguing the pattern predates this release, with the paper and dataset linked in the thread. Useful as the before-picture when the naming conversation comes up.
- Hacker News
Hacker News · thread
Reverse-engineered Jev-like model
An independent reconstruction of the architecture. Read it next to the vendor's quickstart, because the two disagree in places worth knowing about.
- GitHub
GitHub · project
rmalde/minecraft-agent
A planner picks the goal and Jev picks the action, and then the repository publishes the run: nether-final-08 finished in 8 minutes 43.300 seconds against 14 minutes for the previous attempt, with the routes and a read-only sensor kept separate from the game. End-to-end times for a Jev loop are rare enough to be worth the row.
-
Hacker News · thread
Show HN: jevals – replacing LLM judges with typed Jev decisions
Replaces the LLM judge with a typed decision, which is the use case we expected to appear first and the one with the clearest failure mode. The thread's own objection is the one to keep: what happens when the judgement depends on context the typed decision cannot carry.
- GitHub
GitHub · project
TheoLeeCJ/SemIf
Semantic conditionals, run on open models on a single 3090. This is the closest thing in the list to a drop-in replacement for an if-statement, which is the plainest statement of what the idea is for.
- GitHub
GitHub · project
TianyuCodings/NanoJev
A small replica of the mechanics: parallel decisions, dynamic routing. Useful as the readable version of the architecture if you are not going to work through the paper.
- GitHub
GitHub · project
wfzyx/von
The open substitute, published with its own comparison against the model it substitutes: 72.0% macro on a 49-task suite against TypeSafe's 96.6%, and 9.00 kills against 5.62 in ViZDoom — a clear loss on accuracy and a win on the real-time loop. Its GitHub description says sub-15ms and its README headline says sub-25ms; both numbers are the publisher's.
- X
X · video
拿 Jev 做搜索重排，我先泼一盆冷水：单独用，它没打赢向量检索 TypeSafe 的 Jev 这阵子很火，一堆项目拿它做重排。我们在 Agent Skills Hub 的 33,047…
A negative result worth keeping: on their reranking test, Jev alone did not beat vector search. The only post in this column reporting that something did not work.