Colibrì Runs 1.5-TB GLM-5.2 on 25GB RAM, but Crawls at 0.1 Tokens per Second
Updated
Updated · Tom's Hardware · Jul 11
Colibrì Runs 1.5-TB GLM-5.2 on 25GB RAM, but Crawls at 0.1 Tokens per Second
1 articles · Updated · Tom's Hardware · Jul 11
Summary
Italian engineer Vincenzo’s Colibrì proof-of-concept runs the 744-billion-parameter GLM-5.2 model on a modest CPU with just 25GB of RAM and a 1GB/s virtual NVMe drive.
0.05 to 0.1 tokens per second is the trade-off: Colibrì exploits GLM-5.2’s mixture-of-experts design by loading and unloading only the experts needed for each token instead of keeping the full 1.5TB model resident.
INT4 quantization and a lightweight C-based expert selector help shrink and manage the workload, but storage I/O is the first major bottleneck, followed by RAM bandwidth and then CPU core limits.
Colibrì does not yet run on GPUs, and the report says moving expert data to and from a GPU would likely remain the main constraint even with acceleration.
The project is still early and impractical for real-time chat—well below the roughly 20 to 30 tokens per second needed—but it points to a possible path for frontier-scale local AI on high-end consumer hardware.