Case wall
Jev AI Training cases
13 public posts filed under Training. English summaries on this page, with a link to each case.
Last updated
These are the public posts filed under the Training category.
No posts match these filters.
@Matthewagi trained TERN-35B-A3B to try decision models. On an H100 PCIe it reached $0.0116/Mtok, then $0.00773/Mtok on the cost-score Pareto frontier.
The goal of @typesafeai Jev model is to unlock Jevon's paradox. I wanted to play around with decision models so I trained TERN-35B-A3B. The goal is price fronti…
@UnslothAI released a free notebook to train your own decision model so Qwen3.5-4B answers decisions instead of text on 8GB VRAM locally.
You can now train your own Decision model with our free notebook! 💡 Qwen3.5-4B will generate decisions instead of text on just 8GB VRAM locally. Learn to data p…
@eeuhn fine-tuned a legal decision model with Unsloth’s notebook to classify trademarks on the Abercrombie spectrum, hitting 75.8% on LegalBench.
I'm like legal Jev pilled after using Unsloth's train-your-own-Jev model notebook. I made a legal Jev to classify trademarks under the Abercrombie spectrum. Tra…
@nutlope trained Tev1 0.8B, a tiny Jev-like classifier, and shows it classifying tasks locally on a Mac with Ollama at about 50ms end to end. Weights and benchmarks are said to be coming soon.
Just trained Tev1 0.8B, a tiny Jev-like classifier. Here it is running completely locally on my mac with @ollama & classifying some tasks. It's extremely fast:…
@DanRaeder says that while Jev is down for new invites, they worked with Claude to build Vej, a Jev clone that runs locally. They call it weakest at score types and almost 1:1 for noul and choice, and say you can use it to build and play with Jev-based apps until you can get in. The repo is github.com/draeder/Vej.
While Jev is down for new invites, I worked with Claude last night to build Vej (a jev clone that runs locally.) It's weakest at score types, but almost 1:1 for…
@ekzhang1 published a Jev-compatible API that runs Qwen3.6-35B-A3B with SGLang radix cache so prefill can be reused. They say 64 tasks finished in under 1 second.
Inspired by @typesafeai , here is a Jev-compatible public API to play with It runs a comparable open model (Qwen3.6-35B-A3B), and just uses SGLang radix cache t…
@hr98w demos a local visual stand-in for Jev: a 4-bit multimodal Qwen-3.5-0.8B maps candidates to fixed labels, samples after prefill, and limits output tokens. It runs on a 16GB M4 MacBook and sorts three-color goods in about 200ms. They say this copies the form of Jev at inference time, not its real method.
视觉版 Jev 来了 https://t.co/x0kzeE3hdx 仿照着社区的思路，把候选结果映射成固定标签，prefill 完直接采样，限制输出 token，采用多模态的 qwen-3.5-0.8B 4bit 量化部署在本地，16g m4 mbp 顺利运行 叠甲：这只是在 infer 层小小的复现一下 jev 的…
@picocreator introduces simplejev.ai, an open-source library that takes a Hugging Face model and exposes a Jev-like API, including vision, which they say hosted Jev lacks. They say it is on GitHub and in production at FeatherlessAI.
love jev, but upset it - isn't open source? - it lack vision capability? We fixed all of that, introducing https://t.co/DlNeKo2k5e A fully open source library w…
@iamMrDuncan built an alternative of the TypeSafe-style API on Cerebras with Qwen 3.8 27B. They say quality and speed were similar, but TypeSafe was much cheaper and beat Qwen on performance. Source is linked.
Built an alternative version of @typesafeai but on @cerebras with Qwen 3.8 27b. Similar quality, similar performance, but vastly different cost. TypeSafe was wa…
Quoted @CompleteSkeptic
@yuntiandeng describes ProgramAsWeights: small neural programs for fuzzy judgments, compiled from English descriptions, combined with code for control flow. One example joins 30 programs and a decision tree into a course-website helper that looks like a chatbot but runs locally, including a question such as which specialist should receive the query.
I increasingly think future software will combine code for control flow with small neural programs for "fuzzy" judgments. That's what I've been exploring with P…
Quoted @CompleteSkeptic
@theoleecj points people still on the Jev waitlist to an alternative they call close enough, with a browser demo. The post does not name the model behind that demo.
On the Jev waitlist? You can use not Jev (but close enough) right here: https://t.co/S0EQnHkSGL https://t.co/qLtRs0fHyD browser demo https://t.co/BUFN88boT4
While waiting for Jev access, @SimEdw tried the Doom-style demo with Gemma 4 E2B. It takes 640×480 frames directly, shares the image prefill across batched action queries, and reads controls from logits. They say it sort of works.
The Jev Doom demo by @CompleteSkeptic is really cool, so while I’m waiting for access I tried the same thing with Gemma 4 E2B from @GoogleGemma. Mine takes 640×…