Fine tuning with Qwen3.8-Flash-Next

Experiment that edits the model's n-gram embedding table to inject removable facts without fine-tuning.

Tool to add knowledge through the PLE n-gram table.

A lab used a Dell B300 for two fine-tunes of Flash-Next, pruning experts from 512 to 288 per layer with REAP and retraining at NVFP4, GGUF included.

Transferred the pretrained n-gram table into a tiny Qwen model and got 5% lower validation perplexity, trained mostly on free Kaggle GPUs. A resourceful research experiment.

Announced by the builder at UkisAI: a Qwen-based family trained to avoid pathological overthinking, with Flash-Next and 27B variants and 350k+ downloads in 13 days.

Non-uniform quant repo that Strata users commonly pair with the engine.