lmz

Smaller checkpoints. Byte for byte.

Lossless compression built for model weights. zstd -1 takes 22.7% off a Llama-3.1-8B BF16 checkpoint. lmz takes 34.7% — and 64.6% off the directory as Hugging Face actually ships it, which is 13 GB more than zstd on one 8B model.

Nothing is approximated. Every byte comes back.

pip install lmzip
lmz compress ./Llama-3.1-8B-Instruct/

What it saves

Real checkpoints, every round-trip verified byte-identical.

On BF16 weights lmz beats the published state of the art, and sits 0.3 points off the bound no lossless coder of any kind can pass:

Three things a general compressor structurally cannot do, which is where most of the margin comes from: store a tensor once when a directory ships it twice, code a checkpoint as the difference from the one before it, and split on a float's own bit-fields instead of byte boundaries.

Where it is not worth it

Stated plainly, because a compressor that only advertises its wins should not be believed:

Reading a compressed file transparently can never beat a plain one by more than 1/(1−saved) — you still have to read the archive. That is 1.5× on BF16 and 1.05× on Q4_K, so on a fast SSD the mount is slower. It buys disk, not speed.

Also included

lmz add ./my-model/ && lmz mount ~/models   # read a compressed model as ordinary
                                            #   files; llama.cpp needs no patch
lmz fs ~/.lmz/data ~/data                   # a read-write compressed filesystem;
                                            #   32.1% where btrfs+zstd gets 18.9%

Buy me a coffee

lmz is free, MIT-licensed and unfunded. If it saved you disk or bandwidth —

☕ Buy me a coffee

or Alipay (打开支付宝，扫一扫). Thank you.

Documentation

• How it works — why a float array defeats a general-purpose compressor, and the bit-level choices that close the gap
• Measured results — every number, with its conditions
• Using lmz — command line, Python API, the mount and the filesystem
• Limitations — where it does not pay, and what the 81 tests check

Python 3.10+, no runtime dependencies. zstd comes from the standard library on 3.14+; a C compiler, if present, is used once to build the SIMD kernel into the package directory — nothing is installed system-wide. Runs straight from a checkout with ./lmz-cli if you would rather not install it at all.

Check what is active with lmz doctor.