InfoQ: Jev returns typed probabilities

Usable. The piece points readers at pinning a specific model version and comparing other models through the System One adapter, which is integration work rather than a waitlist. The captured page is headline and navigation only, so the specifics here come from the entry copy.

Vendor-reported. The 70 to 500 millisecond range and the low per-million input token price with free output are quoted from TypeSafe, and InfoQ restates them without independent measurement. The entry copy's caveat applies to every figure in the brief.

The two published numbers disagree and the brief only gives one. InfoQ states a 32,000-token window; TypeSafe's model card describes 64k made up of state plus the longest question. The entry copy flags the mismatch rather than picking a winner.

The entry copy describes a single parallel pass over Choice, Score and Noul rather than three sequential calls, billed on input with output free. That is the structural claim to test first, since the price only beats a chat model if you were not paying for generated tokens anyway.

Daniel Curtis, dated 1 October 2026, and the page itself says it is a three-minute read. That length matters for calibration: it is a news brief pointing at TypeSafe's own documentation, not an independent evaluation with its own test harness.

Reinforcement Learning for Calibrated Decisions, abbreviated RLCD, and the piece describes the model as giving up text generation in order to return typed probabilities from a parallel pass over Choice, Score and Noul questions. Anyone evaluating that claim has to go to TypeSafe's own materials for it, since the brief carries no counter-measurement.

Three. Pin jev-1.13.0 rather than tracking a moving alias, read the jaggedness page the brief points at for the rough edges, and compare other models through the System One adapter. That last one is the interesting one for readers: it suggests the brief expects you to want to run Jev alongside other models rather than adopt it exclusively.