What changed
Hugging Face announced the release of Lumma-Fev-27B-v2, the second generation of its Lumma-Fev decision models. This model is part of a family that includes sizes ranging from 0.1B to 27B parameters. The new version scores higher than TypeSafe Jev 1.13.0 across eight public benchmarks and is better calibrated. The model uses bidirectional attention in full-attention layers and causal attention in Gated DeltaNet layers, requiring version 0.3 or later of the checkpoint.
What this means in practice
Lumma-Fev-27B-v2 is designed to take a state and typed questions, returning probabilities for each option in one forward pass without generating text. This allows for direct use in automation workflows where decisions are made based on probability distributions. The model supports multiple languages including English, Hindi, Tamil, and Kannada, with performance metrics showing strong results in both English and Indic languages. The model is trained on 300K samples, up from 100K in the previous version, and uses the decisionformer library for fine-tuning.
Limitations and availability
The Lumma-Fev-27B-v2 model is available for use, with the 0.1B and 0.6B models set to release soon. The model's interface remains consistent with the previous version, allowing existing users to continue without changes. The model's performance is benchmarked against TypeSafe Jev 1.13.0, with Lumma-Fev-27B-v2 showing superior calibration and accuracy across multiple metrics. The model is accessible through Hugging Face's platform, and users can find the models at the provided links.
