Qwen and Jev: local decision models compared

The video compares hosted Jev with independent projects that bring its decision-model approach to local AI. Jev returns answers to supplied questions that software can act on. The speaker stresses that the local implementations are inspired by Jev, rather than downloadable copies of Type-Safe's hosted model.

Nano Jev uses Qwen 3's 0.6 billion parameter model for game controls. In the developer's basic aiming test, it succeeds in all 128 cases, but manages 27 of 128 when predicting a moving target. It receives structured observations rather than interpreting the raw Doom screen. The browser flight-search demonstration has a different limitation: its roughly 7.1-second run uses hosted Jev and Mercury 2.5, and stops at search results.

For private text sorting, the speaker recommends Local Jev with Qwen 3.5 4B as a starting point. It runs offline after downloading the model. Semi's published RTX 3090 test answers 21 binary questions in about one second, versus 5.3 seconds for generated answers. The methods agree on 18 choices, so the timing comparison does not establish identical output.

Hardware and accuracy affect the choice. Local Jev's cited evaluation uses an M4 Max with 36 GB of memory; its timings are not interchangeable with Semi's. The speaker also considers image and audio projects, while noting that Jev Omni samples 16 video frames and caps audio at 30 seconds. A generative model still handles writing and explanation, and consequential actions need separate checks.