Sam Witteveen · 2026-09-20 · notable
Sam Witteveen — 'Open Jev Models Are Here!!'
Sam Witteveen's 20 September 2026 video turns to the open Jev models. OpenJev is an MIT-licensed cross-encoder built on Qwen3.5 that reads a premise and a hypothesis and answers entailment, contradiction or neutral instead of writing text.
Sam Witteveen's channel turns to the open answer to Jev — MIT-licensed models that classify instead of writing text.
What is it?
'Open Jev Models Are Here!!' went up on Sam Witteveen's channel on 20 September 2026. Its subject is OpenJev, an open-weight take on the Jev idea published by AlexWortega on Hugging Face under the MIT license. OpenJev is a single cross-encoder built on Qwen3.5: it reads a premise and a hypothesis and answers with entailment, contradiction or neutral.
How does it work?
OpenJev ships in three checkpoints — a recommended 4B v2 that also reads images, a text-only 4B v1, and a 35B mixture-of-experts variant with 3B active. Because it scores a claim instead of generating prose, the same model reranks answers, grades responses and guards content zero-shot. Its card reports v1-to-v2 gains of 0.42 to 0.63 on ANLI r3, 0.63 to 0.77 on WANLI, 0.52 to 0.84 on image claims and 0.59 to 0.72 on ARC-Challenge.
Why does it matter?
Jev, the System One Model from TypeSafe AI that Witteveen covered two days earlier, is a paid early-access API. OpenJev puts the same shape of model — typed answers rather than text — on Hugging Face under MIT, so the weights can be downloaded and run locally. For engineers who use an LLM mostly to classify and route, that is the difference between a vendor demo and something you can put in a pipeline.
Who is it for?
engineers who classify or route with LLMs