Show HN: A 6M-token movable window on a single 46GB GPU

Other: AI and Machine Learning(arxiv.org)view on HackerNews
AIMachine LearningLanguage ModelsVerified SolutionsFrozen ModelsFrontier ModelsPerformanceAccuracyTokens.

Author: Wetime

Date: 7/28/2026

Article Summary:
The paper presents a frozen 12B language model that beats frontier models on verified work, achieving 100% accuracy and zero tokens, with a persistent memory of verified solutions.