What This Is
We noticed that over the past week, four small open-source models specialized in "judgment tasks" — formally known as Decision Models — have emerged from the open-source community: Laya, Liquid's d1, Cloudflare's Clef-Flash, and Interfaze's Lev. They were benchmarked against each other on a single RTX 4090 GPU. "Judgment tasks" means answering yes/no or belongs-to/doesn't-belong-to questions about a passage of text — for example, extracting amounts from contracts or determining the urgency of a support ticket. These are the most common AI jobs inside enterprises. The test task: read a 9,534-word centipede Wikipedia article word by word and tag every centipede's scientific name.
Industry View
The testers found speed and accuracy don't go hand in hand: Laya takes 3.9 milliseconds per judgment at 97.4% accuracy; Lev achieves the highest accuracy at 98.9%, but is 13 times slower (51 milliseconds). d1 sits in the middle. Clef-Flash comes in second-to-last on speed because its larger model size strains VRAM.
But the objections deserve airtime: the original authors themselves admit the "95%+" accuracy figures are misleading — models answering "no" also count as correct, so the numbers look prettier than they actually are. Another layer of skepticism: a single card running well doesn't equal enterprise-ready. CPU jitter, concurrent request handling, and model fine-tuning costs have not been systematically tested. Reddit commenters also note these models are still closer to "developer toys" than production environments.
Impact on Regular People
For traditional enterprise IT: a single RTX 4090 (around ¥15,000) is enough to run decision-class AI, so the old reflex of "going AI means buying cloud" can be paused — easing both data compliance and cost pressure.
For working professionals: judgment-intensive roles — first-pass customer support, invoice review, compliance screening — will be touched first by these small tools. Not replaced by general-purpose large models, but by specialized small ones.
For the consumer market: running AI locally on phones, appliances, and office devices becomes more viable. Vendors don't necessarily need to ship user data to the cloud anymore; local small models are increasingly sufficient.