July 9, 2026
Hard drive heroics, comment-section chaos
Show HN: Getting GLM 5.2 running on my slow computer
This giant AI somehow runs on a regular PC, and the comments are losing it
TLDR: A developer got a massive chatbot model running on a normal-ish computer by constantly reading pieces from storage instead of needing a giant server. Commenters were impressed, joked about painfully slow speeds, and immediately started worrying whether the trick might chew through SSD life.
A developer just pulled off the kind of stunt that makes the internet sit up straight: they got an absolutely enormous AI model to run on a modest home computer by cleverly keeping only the must-have parts in memory and pulling the rest from storage as needed. Translation for normal humans: this is like fitting a mansion into a studio apartment by only opening one room at a time. The project is tiny, fast to load, and written in plain old C with no extra baggage — and that alone had the Hacker News crowd in full swoon mode.
The loudest reaction was pure admiration. One commenter summed up the mood with the instantly memeable "This is the hacker spirit", while another wondered if this was the spiritual cousin of antirez’s famously scrappy engineering style. But because this is the internet, the applause quickly turned into practical panic: will this absolutely destroy your SSD? That question brought the first whiff of drama, as readers zeroed in on the fact that the system keeps reading huge chunks from disk. Others immediately pivoted into peak tech-chaos fantasy, joking about stuffing the whole thing into a flashy graphics card with built-in storage slots.
And then came the realism brigade. One commenter basically asked the question hanging over the whole demo: is this measured in tokens per second or tokens per geological era? The answer — painfully slow, but maybe useful for overnight jobs — became the thread’s punchline and badge of honor. The vibe was equal parts awe, skepticism, and “this is insane… I love it.”
Key Points
- •The project runs GLM-5.2, a 744B-parameter MoE model, on a consumer machine with about 25 GB of RAM by streaming routed experts from disk instead of loading the whole model into memory.
- •The engine keeps the dense model components resident in RAM at int4 precision using about 9.9 GB, while storing roughly 21,504 routed experts on disk for a total int4 model size of about 370 GB.
- •The implementation is a minimal runtime in pure C with no BLAS, no Python at runtime, and no GPU requirement.
- •Implemented features include validated GLM-5.2 forward inference, MLA attention with compressed KV cache, DeepSeek-V3-style routing, MTP speculative decoding, quantization kernels, async expert readahead, and a native byte-level BPE tokenizer.
- •Reported measurements on a WSL2 system with 12 cores, 25 GB RAM, and NVMe via VHDX show about 30 seconds load time, around 20 GB peak RSS during chat, and cold decode throughput of about 0.05 to 0.1 tokens per second.