March 24, 2026
PlayStation 2? More like ThinkStation 2
Show HN: I ran a language model on a PS2
A PS2 ‘thinks’ from a spinning CD — commenters go feral
TLDR: A developer ran a small chat AI on a PS2 by streaming it off a CD, proving old hardware can do new tricks. Commenters split between loving the retro CD spectacle and pushing for faster storage, while peppering the thread with jokes about “tokens per hour” and questions about speed and quality.
A 26-year-old PlayStation 2 just grew a tiny brain — and yes, it’s booting from a spinning CD. A solo dev ran a small chat-style AI model on a PS2 by streaming the model from disc one chunk at a time, keeping only the essentials in the console’s 32MB of memory. The community? Absolutely buzzing. One camp is cheering this as pure hacker art, with SilentEditor calling it “peak hacker energy,” while others are already hurling questions like, “Okay, but how many tokens per hour?”
The big talking point: that clever CD trick. Commenters like SachitRafa say it sidesteps the tiny RAM the way modern gadgets use flash storage. But the plot twist is speed — with old-school disc read rates, larger models technically “work” but crawl. Cue drama: CD purists love the spectacle of the disc whirring as the PS2 “thinks,” while the pragmatists demand a faster route via the PS2’s hard drive, USB, or memory card adapters. A few hardware heads asked whether those legendary “vector units” could help — the consensus vibe: probably not much.
Amid the applause, there’s real curiosity about latency and quality after quantization (shrinking the model’s numbers to fit). But no matter where you stand, everyone agrees this is ridiculous, brilliant, and deeply nostalgic. Check the demo at naranjositos.tech
Key Points
- •A PlayStation 2 runs transformer inference by streaming model weights from CD-ROM, keeping only activations, KV cache, token embeddings, and RMS norms in 32 MB RAM.
- •The default model is brandon-tiny-10m-instruct (10M parameters, Q8, ~10.4 MB), selected for speed and coherence on PS2 hardware.
- •Larger models can run despite RAM limits; a 77 MB Q4 SmolLM2-135M-Instruct was tested but is too slow on PS2.
- •The project defines a custom PSNT binary weight format supporting ternary, Q4, and Q8 quantization, with provided conversion tools.
- •Build requires the ps2_biw_engine SDK, producing an ELF and bootable ISO; conversion pipelines use Python, NumPy, PyTorch, and scripts for PSNT.