On October 14, FrontiersMind added Lumma-Fev-27B to a family introduced in September. Derived from Qwen3.8-27B, it trains all weights on roughly 100,000 decision examples. Instead of generating free-form text to parse, it assigns probabilities to supplied options, covering English, Hindi, Tamil and Kannada according to the publisher.
FrontiersMind reports 90.7% accuracy on its held-out set of 4,454 decisions, against 83.9% for Jev-1.13 under its protocol. That comparison is not against the Jev-1.14 preview announced in the same week, and the author acknowledges that Jev remains ahead on some reasoning tests. Reported calibration does not guarantee reliable probabilities on a different distribution.
The weights are announced under Apache 2.0; in bfloat16 the publisher quotes roughly 49GiB and recommends an 80GB GPU. The interface follows the /v1/systemone contract, but training and evaluation covered only noul and choice questions. The interface accepts score without validating it in this release. For automation, request compatibility is therefore not decision validation.