The latest articles on DEV Community by Nikola Nikolov (@nikolovv). https://dev.to/nikolovv https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4135892%2F63b8d4da-f84c-4ae1-94fe-feb11e72a6aa.png https://dev.to/nikolovv en Nikola Nikolov Fri, 25 Sep 2026 11:33:48 +0000 https://dev.to/nikolovv/jev-new-paradigm-or-just-logits-with-a-better-marketing-team-5567 https://dev.to/nikolovv/jev-new-paradigm-or-just-logits-with-a-better-marketing-team-5567 <p>Jev is being pitched as a brand new kind of model: give it a fixed set of choices, get calibrated probabilities back in one shot.</p> <p>but isn't that just a forward pass and a softmax? you can already do that fully on-device:</p> <ul> <li> <strong>NobodyWho</strong>: open-source, drop-in local LLMs for basically any stack (Python, Swift, Kotlin, Flutter, RN, Godot)</li> <li> <strong>Cactus</strong>: on-device inference for phones, already routes by model confidence</li> <li> <strong>llama.cpp</strong>: the OG, grab the logits yourself</li> </ul> <p>I get that the secret sauce is supposed to be calibration, and sure, raw logits from a small model aren't calibrated and usually have a favorite letter. but is that a real moat?</p> <p>what am I missing? has anyone actually compared them?</p> ai privacy local webdev