Basal-Omni: An open-weight 31B model for AI decision-making (nr 1 on JevBench) (huggingface.co)

The newly announced basal-omni-1.0-ultra is a state-of-the-art AI decision-making model featuring 31 billion parameters, making it the largest in its series. Unlike its predecessor, basal-1.5, this model not only processes textual data but can also respond to image-based queries using the Gemma 4 image encoder, albeit in a zero-shot capacity. This dual capability enhances its functionality to answer complex questions with calibrated probabilistic outputs, significantly increasing the number of options per query from 10 to an impressive 255. This release is a significant advancement for the AI/ML community, particularly in decision-making contexts where both text and images interact. The model's experimentation with System 1-type decisions allows it to interpret states (like web pages or documents) and provide answers based on pre-defined options in a single forward pass, optimizing speed and efficiency. With an impressive performance benchmark reflecting a macro score of 0.879 in Polish and English contexts, basal-omni-1.0-ultra promises to set new standards in automated decisions, especially in areas requiring nuanced assessments such as contract analysis and document interpretation. Despite its limitations regarding audio processing, the academic and practical implications of this open-weight model could reshape how AI systems handle complex decision-making tasks in the future.