Tag
This paper presents a frozen 12B model that uses a persistent memory of verified solutions to achieve 100% accuracy on solved problem families with zero generation tokens, and demonstrates a 6M-token movable window on a single 46GB GPU.
Release of an uncensored version of Gemma4-12B with quantization-aware training (QAT) and a multi-token-prediction head for ~60% faster generation via speculative decoding. The model removes refusals while maintaining quality, and is available on Hugging Face.