Hi — first of all, wonderful project. 🐦

I ran the full pipeline on my machine and published the result, so others can skip the 756 GB FP8 download and the conversion:

https://huggingface.co/jlnsrk/GLM-5.2-colibri-int4

• Produced with the unmodified converter: coli convert (--ebits 4 --io-bits 8), including the MTP head — 141 shards + 3 MTP shards + config/tokenizer files, 378.8 GB total, file sizes verified against my local copy.
• Model card explains it's colibrì's own container (not GGUF/AWQ) and links back to this repo. MIT, matching the base weights.

Notes from the run, in case they're useful datapoints:

• Hardware: WSL2 (Ubuntu 24.04) on Windows 11, 24 cores, 24 GB RAM, model dir on the WSL ext4 VHDX.
• Download + conversion of all 141 shards took ~7 h end-to-end (~3.7 min per shard cycle at ~25 MB/s network), fully unattended with supervisor.sh (adapted paths). The resume logic survived a mid-run reboot flawlessly.
• setup.sh built clean with gcc 13, no warnings that stopped anything.

Feel free to link the mirror in the README if you find it useful — and thanks for building this!