FrontiersMind/Lumma-fev-0.6b

FrontiersMind developed this 649 million parameter system to classify input documents without generating new text. It accepts a state and typed questions, then outputs probability distributions for each query in a single forward pass. The architecture supports English and ten Indian languages, using a pointer readout over predefined options. Developers can load the open weights through the transformers library or the dedicated lumma-fev Python package. The model performs single-pass decision tasks, processing text, JSON objects, or arrays for questions categorized as yes-no, multi-choice, or ordered scores. Users may also serve the model as an API using the lumma-fev-serve command, which exposes an endpoint compatible with existing TypeSafe clients. The creators note that larger versions are releasing on 26 September with a detailed blog. Confidence scores are the model's own estimates, so FrontiersMind advises measuring calibration on labeled data. Run internal tests before gating automated actions on these probabilities, as the model relies on its internal estimates for accuracy.