Remek/basal-omni-1.0-ultra · Hugging Face

Why it matters

Hugging Face hosts the model card for Remek/basal-omni-1.0-ultra, an experimental release described as the largest model in the basal-omni line, built on Google Gemma 4. The model takes a different approach from typical language models: rather than generating text, it reads an input state, such as a message, document, or agent trace, and answers a typed question about it by producing calibrated probabilities for each allowed option in a single forward pass. It was trained on Polish and English typed decisions using soft targets from an earlier basal release. The card highlights several changes over that earlier model, including a larger Gemma-based foundation, far more options per question, and zero-shot image understanding through Gemma's own encoder, though the tier lacks any audio support. Benchmark results show gains over a sibling model on Polish decisions and several English reasoning tasks, with one financial dataset as an exception. The card also documents deployment through libraries, inference servers, and container tooling.