Deel's Jev numbers, posted by TypeSafe
TypeSafe AI published Deel's measurements on six classification tasks, including repeat-question matching going from 70% to 97% and expense categorization from 50% to 86%. Every number here comes from the vendor's post, not an independent run.
Built with it
Hassan's Tev1 0.8B runs on his Mac at about 50ms end to end
Hassan trained Tev1 0.8B, a smaller sibling to the tev1-4B classifier he put out earlier, and filmed it classifying tasks locally through Ollama. He put end to end latency at roughly 50ms and said the video is not sped up. Weights and benchmarks are promised but not out yet.
It's extremely fast: only ~50ms E2E latency. Video is not sped up!
Someone measured it
TypeSafe posts Deel's numbers on six Jev classification tasks
TypeSafe AI published what Deel measured across six pick-from-a-known-set tasks, including matching repeat analytics questions, blocking PII requests and sorting expenses into about 55 categories. Repeat-question matching went from 70% to 97%, expense categorization from 50% to 86% against human reviewers, and escalation caught the same cases with fewer false alarms.
Speed was measured while shadowing live production traffic: up to 4× faster.
Writing criteria
Jack Cheng puts Jev on both ends of an email triage pipeline
Jack Cheng wired Jev into two stages of an email classification system, a shape he calls a Jev sandwich. The second pass decides what deserves his attention based on how well he slept and his current mood, so the same inbox sorts differently on different days. He describes it as an experiment, with no accuracy numbers attached.