Justvugg/GLM-5.3-colibri-int4-g64

GLM-5.3 for colibrì (int4-gs64)

An int4 group-scaled conversion of zai-org/GLM-5.3, packed in the container format that colibrì streams from disk.

This is not the official checkpoint. The weights are Z.ai's; this repository only changes how they are stored. It will not load in transformers or vLLM. It loads in colibrì's colibri engine and nothing else.

What it is for

Running a 744B-class model on a machine with about 25 GB of RAM, by keeping the routed experts on disk and reading only the ones each token selects.

on disk	419.3 GB across 141 shards
resident dense	10.6 GB
total RSS	about 13.6 GB with an 18 GB budget
shape	78 layers (75 sparse), 256 experts, 8 active per token
routed experts	int4, group size 64, 21.2 MB each
dense / io	8-bit
MTP	not included in this conversion

Use

git clone https://github.com/JustVugg/colibri && cd colibri
make -C c colibri
c/coli chat --model /path/to/this/folder

The engine detects the family from config.json and picks colibri automatically.

One thing the output gets wrong today

Z.ai state it plainly on the upstream card:

GLM-5.3 uses the same base model as GLM-5.2. Every gain comes from post-training.

That is visible in the files. The two checkpoints have identical geometry, and their config.json differ only by the transformers_version that wrote them and one extra key. There is nothing architectural to tell them apart.

The practical consequence is that colibrì currently announces this checkpoint as "GLM-5.2 744B", because that is what the configuration says and no heuristic over the configuration can ever say otherwise. The engine loads and runs the right weights; only the label is wrong. Naming it correctly requires taking the name from outside the config, and that is tracked in the colibrì repository.

Provenance and what has been verified

Converted from the official zai-org/GLM-5.3 release with colibrì's converter at ebits=4, xbits=4, io_bits=8, group_size=64.

All 141 shards were checked for header validity and truncation before upload: 116,915 tensors, no problems found.

What has been confirmed by running it: the model loads, generates coherent text, and produces byte-identical output across two different expert-fetch paths under a fixed seed. What has not been done: a token-exact comparison against the reference implementation. If you run that comparison, please open an issue in the colibrì repository with the result, whichever way it goes.