Colibri Demonstrates Large-Model Inference on 25GB of RAM With Heavy Speed Trade-Offs

Tom's Hardware reports that the Colibri project has been used to run the 744-billion-parameter GLM-5.2 model, described as roughly a 1.5TB model, on a modest CPU system with 25GB of RAM and a 1GB/s virtual NVMe drive. The approach relies on the model's mixture-of-experts structure, repeatedly loading only the needed expert slices instead of keeping the full model resident in high-end accelerator memory. The demonstration comes with a major limitation: throughput is reported at only about 0.05 to 0.1 tokens per second, making it impractical for real-time interaction. Even so, the project is relevant to memory watchers because it underlines the trade-offs between storage speed, memory capacity, memory bandwidth, and compute resources in AI inference. For the memory market, the news does not directly change DRAM pricing or supply, but it reinforces the broader theme that AI workloads are increasingly shaped by how efficiently systems can use limited RAM and fast storage.