Sebastian Raschka puts Jev in the context of text classification history

Sebastian Raschka published a long write-up tracing text classification from bag-of-words to TypeSafe's Jev, and where Jev fits.

Sebastian Raschka has published a long write-up (Ahead of AI, September 29, 2026) placing TypeSafe's Jev in the history of text classification. For anyone doing triage or routing, his framing is the useful part. Frontier LLMs can do the same classification tasks, but Jev's advantage is speed and cost. For a narrow, well-defined problem, he says Jev "probably won't" beat a special-purpose classifier. Jev's case is generality. His own view moved from "I can easily build this myself" to "this actually works better than I thought." He says he is not affiliated with TypeSafe and did not get free access.

Much of the piece is a practical baseline refresher. On the IMDb review dataset, he cites about 89.9% accuracy for bag-of-words with logistic regression and 85.66% for an LSTM trained from scratch. ULMFiT reached 95.4% with pre-training. Bag-of-words plus logistic regression remains his go-to first baseline.

He describes his account of Jev's methodology as an educated guess. TypeSafe has not published its architecture. The vendor's launch post says Jev is in early access and claims calibrated probabilities at $0.042 per million input tokens. It also says it can't prove that price isn't subsidised. Its speed figures are its own benchmarks. Independent tests so far are task-specific. Towards Data Science warns against treating early numbers as universal. Cloudflare lists Jev on Workers AI if you want to test it against your own labelled data.

Sources