← back to projects
ZipLLM — efficient LLM storage
Model-aware deduplication + BitX delta compression to store many LLM checkpoints cheaply.
The idea — dedup + delta compression
Base model
Fine-tune A
Fine-tune B
T0 (shared)
T1 (shared)
T2 (shared)
T3 (shared)
1× stored
T0 (shared)
T1 (unique)
T2 (shared)
T3 (unique)
T0 (shared)
T1 (shared)
T2 (unique)
T3 (unique)
⊕
⊕
⊕
⊕
BitX delta
BitX delta
BitX delta
BitX delta
Step 1 — deduplication: identical tensors across checkpoints collapse to one shared copy.
Step 2 — BitX delta compression: each unique tensor is XOR-ed (⊕) against the base tensor,
and the resulting small delta is losslessly compressed.
The result — storage shrinks by ~54%
Shared / deduplicated tensor
Unique tensor (delta-compressed)
BitX compressed delta
Saved space