Sebastian

Gemma 4 12B Unified (encoder-free audio): audio comprehension collapses past ~450–500 tokens of absolute position — vision on the same checkpoint is unaffected, E2B/E4B (Conformer) are unaffected