Show HN: A 6M-token movable window on a single 46GB GPU
Category: ai-ml
Tags: frozen-model, verified-memory, long-context, zero-token, exact-retrieval
Score: 6.7/10 (Innovation: 7, Technical: 7, Documentation: 5, Utility: 6)
This project proposes a frozen language model paired with a persistent memory of verified solutions, achieving zero-token, bit-exact answers for previously solved problem families. It also demonstrates a 6M-token movable context window on a single 46GB GPU, far exceeding vLLM and SGLang limits, by using exact addressing instead of approximate similarity retrieval.
Target audience: AI researchers, ML engineers, backend developers working on inference optimization
Repository: https://arxiv.org/abs/2607.23806
View on Hacker News