Modern AI on Vintage Hardware: Llama 2 Runs on Windows 98 (hackaday.com)

EXO Labs rebuilt and ported a version of Llama 2 to run on a Windows 98-era PC, showing a Pentium II machine with 128 MB of RAM generating ~39.31 tokens/sec with a 260K-parameter model and ~1.03 tokens/sec with a 15M-parameter model. Getting it to work required classic-hardware hacks—PS/2 peripherals, FTP-over-Ethernet file transfers and some source-level porting/compilation—but the demo proves even very small LLMs can be executed on decades-old hardware. Bigger models will run too, but with progressively slower token rates; a key takeaway is practical feasibility rather than parity with billion-parameter models. The broader significance is about resilience and decentralization: if modern AI can be made to run on Windows 98, it’s harder for large companies to fully control how models are used and distributed. EXO Labs frames this as one of a dozen projects aimed at making AI runnable “anywhere, by anyone,” and points to existing single-file, cross-platform command-line executables and GitHub toolchains for local AI. Technically, the demo underlines trade-offs—model size vs. latency and utility—but also the surprising power of narrow, efficiently trained small models for real tasks, reinforcing the movement toward self-hosted, offline, and hardware-agnostic AI.