A 744B AI Model Runs On A Laptop (GLM-5.2 Via Colibri)
GLM 5.2 runs local on a laptop via Colibri's disk-streaming trick, no GPU, no H200 cluster needed GLM 5.2, Zhipu AI's 744-billion-parameter open-weight model, is now running on a normal 12-core laptop with no GPU thanks to a free tool called Colibri, and the reason why says more about AI news in general than about this one model. GLM 5.2 is a Mixture of Experts model, meaning only about 40 billion of its 744 billion parameters activate per token, and of those, just ~11GB actually change from token to token. Colibri, written by a solo developer (JustVugg) in about 1,300 lines of pure C with zero dependencies, keeps a ~9.9GB int4 dense core resident in RAM and streams the other 370GB of routed experts straight off an NVMe SSD on demand. Cold-cache speed is brutal, 0.05 to 0.1 tokens/second, but native multi-token prediction and a warm cache push that to 2.2-2.8 verified tokens per pass, and on an Apple M5 Max testers hit roughly 1.06 tokens/second, enough to make it genuinely usable. GLM 5.2 itself scores 51 on the Artificial Analysis Intelligence Index, the top mark for any open-weight model, and beats GPT-5.5 on Terminal-Bench 2.1 coding tasks at a fraction of the API cost. This video breaks down the MoE architecture that makes this possible, why running frontier weights no longer requires a datacenter, and what tradeoffs (speed for memory) actually look like in practice. It's for builders running local LLMs, anyone curious about open webui setups, deepseek free api alternatives, or the broader shift toward local AI content creation, image to 3d and text to 3d workflows, and artificial intelligence tools that don't need a cloud bill. Chapters: 0:00 Intro 0:21 The Lie Behind The Parameter Count 1:23 Why You Don't Need A Datacenter 2:38 Inside The 744B Giant's Real Trick 3:36 The Engine That Streams From Disk 4:30 The Brutal Cost Of Going Slow 5:33 Guessing The Future To Get Faster 6:35 Just How Smart Is This Model 7:58 Why This Is Winning The Internet Tools & resources mentioned: - Colibri (GitHub): https://github.com/JustVugg/colibri - GLM-5.2 (Hugging Face / Z.ai) - Open WebUI: https://github.com/open-webui/open-webui - Hacker News: Show HN post: https://news.ycombinator.com/item?id=1 About The Stack The Stack helps you build with AI. Each video takes one tool, model, or workflow and shows how it works in a few focused minutes, with the real benchmarks and real costs. We go deep on Claude Code and Cursor for AI coding, AI agents and MCP servers, the open-source AI tools and GitHub repos most people miss, RAG and vector search, fine-tuning, and running local LLMs on your own machine with Ollama and LM Studio. We compare models like ChatGPT and Claude, test AI automation with Zapier, Make, and n8n, and flag the tools that actually ship. Subscribe for new breakdowns: https://www.youtube.com/@the-stack-ai?sub_confirmation=1 #glm52 #aiagents #localllm #opensourceai #moe
