Research experiments

Qwen3-0.6B PerfectBlend 1024

Fast-iteration DFlash setup: Qwen3-0.6B draft, 3 draft layers, 3 target hidden features, sequence length 1024, 100k regenerated PerfectBlend samples.

draft_model Qwen3-0.6B
target_model Qwen3-8B
draft_layers 3
target_hidden_features 3
max_length 1024
samples 100000
cache_size_gb 787
cache_location /mnt/ramcache (tmpfs)
seed 42

Used by