Open-weight & owned stack · Training tuning
Qwengram Transfers N-Gram Memory Into Qwen
Ninnix96 published checkpoints and training code for a Qwen3.5-0.8B variant with lower validation perplexity after transferring frozen n-gram memory.
Read the original at reddit.comOpens the publisher's site in a new tab