Qwengram-0.8B

Qwengram-0.8B: n-gram memory transfer cuts perplexity 5.05%

A developer transferred Qwen3.8 Flash-Next's pretrained PLE n-gram memory into Qwen3.5-0.8B, keeping the backbone and ~51B-parameter memory frozen while training an R=1 reader; validation perplexity fell from 18.2759 to 17.3534, a 5.05% reduction. This is a language-model validation result, not a 5% benchmark accuracy gain.