Turn GLM-5.3-Flash into a Jev-class decision model, no fine-tuning

The open recipe reads logprobs to hit Jev's accuracy at ~180ms with images Jev lacks — plus open clones and a font that shows how models see your code.

Nowline SEP 27 10:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • No fine-tuning: match Jev by reading GLM-5.3-Flash's logprobs

    A merged token-factory-cookbook recipe numbers your options, prefills `choice_index:`, and reads the probabilities GLM-5.3-Flash assigns at that one position in a single forward pass — no text is generated. It ties Jev across 28 datasets (median 0.7pp gap, p=0.64) at roughly 180ms per decision and ~$0.00008 each, and unlike Jev it can decide over images. Works on any vLLM-backed endpoint.

  • See your prompt the way the model does

    A new free web font renders text so every LLM token takes the same width, making token boundaries visible at a glance. Paste in your code or a system prompt and you can literally see where the tokenizer splits it — and where you're quietly burning tokens.

  • The open Jev-clone weekend

    Don't want a proprietary dependency? The reproductions have landed. Laya, a 421M Apache-2.0 model, returns typed, calibrated yes/no, score, and choice answers in ~33ms self-hosted and can't emit prose or malformed JSON by construction. Jev's own pitch — up to 193x faster and 445x cheaper than a chat LLM for decisions — is now something you can approximate on hardware you own.