I ran a language model on a PS2
Details
- External ID
- 47470405
- Source
- HN
- Company
- —
- Product
- I ran a language model on a PS2
- Website domain
- github.com
- Launched
- March 21, 2026
- Cohort
- —
- Upvotes
- 46
- Upvotes percentile
- 0.8290282902829028
- Tags
- —
- Fetched at
- Sept. 7, 2026, 9:26 p.m.
- Updated at
- Sept. 7, 2026, 9:26 p.m.
Description
The Emotion Engine has 32 MB of RAM total, so the trick is streaming weights from CD-ROM one matrix at a time during the forward pass — only activations, KV cache and embeddings live in RAM. This means models bigger than the RAM can still run, they just read more from disc.Had to build a custom quantized format (PSNT), hack endianness, write a tokenizer pipeline, and most of the PS2 SDK from scratch (releasing that separately). The model itself is also custom — a 10M param Llama-style architecture I trained specifically for this.And it works. On real hardware.
Enrichment
- Theme
- lightweight and on-device AI runtimes
- Vertical
- Horizontal
- Function
- Model & infra
- Audience
- Developer
- AI stance
- AI-native
- Project type
- Hobby / open-source project
- Normalized one-liner
- run language models on legacy hardware
- Manually corrected
- False
Could you build this?
No Porting an LLM forward pass to bare-metal PlayStation 2 hardware (Emotion Engine) with 32 MB of RAM and optical disc streaming requires deep embedded systems, assembly/MIPS architecture, and custom quantization expertise.
What it would actually take: Requires low-level C/MIPS assembly and PS2 homebrew SDKs (PS2DEV). The engineering involves designing a custom quantization pipeline, paging matrix weights synchronously or via DMA from the CD-ROM drive into tight scratchpad/main RAM, and handling hardware vector units (VU0/VU1) for inference math without standard OS abstractions.
Discussion
12 comments analyzed.
Competitors mentioned: Open PS2 Loader, PS2Linux, RenderWare
Concerns raised: PS2 processors can't push full 100Mbit ethernet speed, USB 1.1 much slower than CD interface, 32MB RAM constraint limits model size, Development still hard for newcomers, CD-ROM streaming latency with larger models
Feature requests: Support for larger models via HDD/Memory Card adapters, Leverage EE vector units for inference performance, Multiplayer support for Age of Empires II, Open source RenderWare
Competitors
Other products that read as similar to this one — 110 launches clear the similarity bar, closest 8 shown.
Attention rank: #30 of 111 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).
Launched 143 days after the earliest competitor.
- Z80-μLM, a 'Conversational AI' That Fits in 40KB · hn · 2025-12-29 · 514 upvotes · similarity 0.46
- Moonshine Open-Weights STT models · hn · 2026-02-24 · 316 upvotes · similarity 0.43
- L88 · hn · 2026-02-24 · 12 upvotes · similarity 0.42
- Serve 100 Large AI models on a single GPU with low impact to TTFT · hn · 2025-11-08 · 7 upvotes · similarity 0.41
- Wienerdog · hn · 2026-08-01 · 9 upvotes · similarity 0.40
- Watch 14-Byte AI "brains" attempt to solve a 2D maze (Its hard) · hn · 2026-07-27 · 25 upvotes · similarity 0.39
- I built a tiny LLM to demystify how language models work · hn · 2026-04-06 · 915 upvotes · similarity 0.39
- I run 30B 22tok/s, 109tok/s not novel,6GB/16GB RAM overcoming llama.cpp · hn · 2026-07-29 · 5 upvotes · similarity 0.39
Other launches for this product
- No other launches for this product.
Same idea, different domain
Nobody's really built a model & infra tool for Fintech yet.