HuggingFace released Lumma-Fev-27B on October 14, 2026, saying it scores options in a single forward pass instead of generating a label. It is the company's largest decision model to date, following the release of Lumma-Fev with parameters ranging from 154M to 9B.
HuggingFace reported 90.7% accuracy on a held-out set of 4,454 decisions, measured against TypeSafe Jev's 83.9%. That compares with Jev's performance on the same rows, with Lumma-Fev-27B showing a +6.9 point improvement across nine source tasks.
Lumma-Fev-27B is built on Qwen3.8-27B and targets decision-making in English and three Indic languages: Hindi, Tamil and Kannada. Availability begins with the model's release, initially for developers and researchers.
"It is the first Fev model we compare head-to-head with TypeSafe Jev on the tasks the family was built for, and we report where it wins and where it does not," said FrontiersMind. The model is trained and evaluated on noul and choice questions only.
The announcement follows the release of Lumma-Fev, a family of decision models. HuggingFace did not say how the model will perform on reasoning benchmarks such as BBH and WinoGrande, where Jev is still ahead. The company said it will continue to refine the model based on user feedback.
Source: huggingface
