🚀 Holy 💩! Major Local AI Breakthrough! 🧠 744B-parameter GLM-5.2 (1.5 TB total) is now running on just ~25 GB RAM — no discrete GPU required! 👀 Wut!? Must test!! Italian engineer @JustVugg built Colibrì, a pure C inference engine (single ~2.4k line file, zero runtime deps) that: • 🛡️ Keeps the dense core (~10 GB at int4) resident in RAM • 📀 Streams 21,504+ MoE experts from fast NVMe on demand (only ~40B active per token) • ⚡ Supports native MTP speculative decoding + MLA attention Result: Frontier-class model on everyday consumer hardware! 📊 Current speeds: • 25 GB RAM setup → 0.05–0.1 tok/s (disk-bound) • Higher RAM + fast SSD → up to 1+ tok/s (warm) • It's a start... what could you do with a 5090? 💡 Big opportunity: Pair it with Phison aiDAPTIV+ AI SSDs to kill the I/O bottleneck 👉 smarter caching, prefetching & KV offload could make it dramatically faster! This is a huge step toward truly accessible local frontier AI. 🔗 GitHub: JustVugg/colibri