Show HN: Getting GLM 5.2 running on my slow computer

	Show HN: Getting GLM 5.2 running on my slow computer(github.com)
	922 points by vforno 63 days ago \| 238 comments
	tl;dr: Colibrì is a dependency-free C engine that runs GLM-5.2 (744B-parameter MoE) on modest hardware (~25GB RAM, 12 cores) by keeping the ~17B dense parameters resident in int4 while streaming the 21,504 routed experts from disk (~370GB) via an LRU cache. It implements MLA attention, MTP speculative decoding, DSA sparse attention, and an OpenAI-compatible API, achieving ~0.05–0.1 tok/s on the author's WSL2 dev box, with community benchmarks reaching ~2 tok/s on an M5 Max and ~1 tok/s on a 430GB EPYC system. Performance scales with RAM (cache size), disk bandwidth, and matmul throughput.
	HN Discussion: ↓The README appears AI-generated based on repeated use of 'honest' phrasing ↓The reported tok/s is too slow to be practically useful even for overnight tasks ↑Others are working on similar streaming/LRU weight approaches and share their parallel efforts ↑This points toward a future where fast SSDs replace expensive RAM for LLM inference ~Questions whether this offers advantages over llama.cpp's existing mmap and quantization support