Caltech-backed PrismML released Bonasi, a 1-bit LLM that is 14x smaller and 5x more energy efficient than comparable 8B models, aiming to run AI without cloud dependency.