Sam Witteveen · 2026-09-21 · notable
Sam Witteveen — 'How to Build Things with Jev & OpenJevs'
Sam Witteveen's 21 September 2026 video is a build-along with Jev and the OpenJev models. OpenJev is an MIT-licensed cross-encoder finetuned from Qwen3.5-4B that returns entailment, contradiction or neutral instead of writing text.
Sam Witteveen follows his two Jev episodes with a hands-on build using Jev and the open OpenJev checkpoints.
What is it?
'How to Build Things with Jev & OpenJevs' went up on Sam Witteveen's channel on 21 September 2026, his third Jev episode in four days. Its subject is OpenJev, an MIT-licensed cross-encoder finetuned from Qwen3.5-4B that reads a premise and a hypothesis and answers entailment, contradiction or neutral rather than writing prose.
How does it work?
Four OpenJev checkpoints are published: a 0.8B text model with a 4k context, a recommended 4B v2 that also reads images, the original text-only 4B, and a 35B mixture-of-experts variant with 3B active. Because the model scores a claim instead of generating text, one checkpoint covers reranking answers, grading content and agent reasoning. Its card reports 0.91 on MNLI and 0.63 on ANLI r3 for the 4B v2.
Why does it matter?
Sam Witteveen's channel is where working engineers go for build-alongs, and this episode moves Jev from 'here is a new kind of model' to wiring one into something. For anyone whose LLM calls end in a parsed yes/no, OpenJev's MIT licence and published weights mean the swap can be tried locally, at four different sizes, without an API key.
Who is it for?
engineers building classification and routing steps