Show HN: Getting GLM 5.2 running on my slow computer

	Show HN: Getting GLM 5.2 running on my slow computer(github.com)
	729 points by vforno 1 day ago \| 179 comments
	tl;dr: A developer built "colibrì," a dependency-free C engine that runs the 744B-parameter GLM-5.2 MoE model on consumer hardware (~25GB RAM) by keeping the dense weights (~10GB int4) resident and streaming the 370GB of routed experts from NVMe on demand, with an LRU cache and MTP speculative decoding. Performance is disk-bound: ~0.05–0.1 tok/s on the author's WSL2 machine, but community benchmarks show ~1 tok/s on an M5 Max and projections of 5–15 tok/s on beefier hardware. Includes an OpenAI-compatible API and optional CUDA backend for pinned hot experts.
	HN Discussion: ↓Suspicion that the project's writeup is AI-generated based on stylistic tells like overuse of 'honest' ~Questioning practical usability given the extremely low tokens/second performance reported ↑Excitement about similar streaming/offloading approaches and sharing parallel projects ↑Optimism that SSD streaming could replace expensive RAM as a viable architecture for local LLMs ↓Skepticism that this offers real gains over existing solutions like llama.cpp's mmap approach