Solo Developer Runs GLM 5.2 — a 744B MoE Model — on a Consumer Laptop with 25 GB of RAM

Colibrì streams the 744-billion-parameter GLM 5.2 model from disk on consumer hardware, pure C with zero dependencies and no GPU.

A developer working alone has done something that reshapes what "running a frontier model" even means: GLM 5.2, a 744-billion-parameter mixture-of-experts model, now runs on a laptop with just 25 GB of RAM.

The project, Colibrì, is a single ~2,400-line C file. No Python at runtime. No GPU. No BLAS libraries. The insight is simple but radical: only about 40 billion of the 744 billion parameters activate per token, and only 11 GB worth of those change from token to token. So the dense part — attention, shared experts, embeddings — stays resident in RAM at int4 precision, consuming about 9.9 GB. The remaining 21,504 routed experts, about 370 GB at int4, live on disk and stream in on demand, layer by layer, with a per-layer LRU cache.

It is not fast. Cold starts crawl at 0.05 to 0.1 tokens per second, bottlenecked by NVMe random reads. But with multi-token prediction speculative decoding, acceptance rates of 39 to 59 percent push throughput to 2.2 to 2.8 tokens per forward pass once the cache warms up. The model uses DeepSeek-style MLA attention with a compressed KV-cache that is 57 times smaller than the uncompressed version, plus DSA sparse attention and KV-cache persistence across restarts — close the terminal and reopen it warm.

The developer, who goes by JustVugg, wrote and tested everything on a 12-core laptop with no access to server hardware. The project appeared as a Show HN on July 9 and drew 882 points and over 200 comments from developers who saw the implications immediately: when a $1,500 laptop can run a frontier-class model that competes with Claude and GPT, the economic floor for inference has dropped through it.

A pre-converted int4 model is available on Hugging Face. The engine is a single CLI binary. Load time: 30 seconds. Frontier AI, once the exclusive province of GPU clusters worth millions, now answers from a consumer NVMe drive.

Sources: GitHub: JustVugg/colibri · Hacker News Show HN