Run a 700-billion-parameter GLM on a laptop: Colibrì turns SSD into VRAM and trends on GitHub

An open-source project called Colibrì is climbing GitHub fast. According to a report by QbitAI, it offers a practical, hands-on route to running a 700-billion-parameter GLM model on a laptop — without a GPU.

The trick is straightforward. Instead of holding a 744B model entirely in memory, Colibrì keeps only about 9.9GB resident and parks the rest on the SSD, effectively using solid-state storage as VRAM.

That pulls the storage hierarchy back into the inference path. Normally a model has to fit in VRAM or RAM, and once weights vastly exceed available memory, inference simply cannot start. Treating the SSD as a weight store trades bandwidth for capacity and lowers the entry barrier.

The trade-off is obvious. SSD bandwidth and latency are in a different league from VRAM, and the source emphasises ease of getting started and feasibility rather than publishing throughput numbers. Real-world experience will depend heavily on local drive performance and access patterns, so usability still needs community measurement.

For the wider ecosystem, projects like this push out the boundary of local inference: lower hardware requirements, better data privacy, and a testbed that does not depend on cloud compute. When very large models can only run in data centres, individuals and small teams struggle to take part in validation; putting weights on local storage at least makes getting something running possible.

What to watch next is community measurement — real speeds on different hardware, support for more open models, and whether caching and prefetching can move the SSD approach closer to usable. For open-source work, staying maintained after the initial spike matters more than a one-night trend.