Can a Laptop Really Run a 744 Billion Parameter Model?

A developer ran GLM 5.2, a 744 billion parameter open model, on a laptop with 25 GB of RAM and no GPU, using Colibrì, a small open-source engine written in C. It works because GLM 5.2 is a mixture-of-experts model, so only about 40 billion parameters are active for each token. Colibrì keeps about 10 GB of shared weights in RAM and streams the rest, around 370 GB of experts, from the SSD. It works, but slowly: one token every 10 to 20 seconds. Around the same time, Andrea Borio got about 2 tokens per second on a 64 GB MacBook Pro by keeping frequently used experts in memory, using a heavily compressed version of the model. Running local inference for massive models on consumer hardware is usually impossible. I show how to run the 744 billion parameter GLM 5.2 model using only 25 GB of RAM. Large language models typically require hundreds of gigabytes of VRAM to function, making them inaccessible for most developers. By shifting the workload to an SSD instead of a high-end GPU, I demonstrate a method to execute these heavy models on a standard laptop. This approach highlights the trade-offs between memory speed and capacity when you are working with strict RAM constraints. While the processing speed reaches 10-20 seconds per token, this experiment provides a blueprint for those interested in LLM hardware requirements and local model deployment. Understanding how to swap memory for storage is a practical skill for engineers looking to test state-of-the-art architectures without expensive infrastructure. Subscribe for weekly videos on AI engineering, LLM optimization, and building production-ready AI systems! #LocalLLM #GLM #MixtureOfExperts #OpenSourceAI #LocalAI