Engineer runs Gemma 3 270M entirely inside an office laser printer at 60s per token
osanseviero · x · 2026-11-17
Hugging Face engineer osanseviero shows an absurd but real demo: using the Carbon compiler, a small transformer engine plus quantized Gemma 3 270M was compiled into a single PostScript job and executed entirely inside an office laser printer.
- The printer's PostScript interpreter handles all inference, with only 14MB peak RAM
- Inference runs at a blistering 60 seconds per token, using careful layer juggling to avoid OOM
- The printed page itself shows the hardware receipt, tokenizer output, and the model's generated reply: "Read the fine print"
More from Fun
- Someone got DOOM running on a Nikon Z6 II and open-sourced a full Expeed 6 QEMU emulator — jedisct1 · 2026-11-17
- KNOW THYSELF: A documentary on what AI looks like from Athens, coming soon — realmeetjames · 2026-11-17
- GPT-4o generates pixel-art robot journaling sequence with smoothly shifting light — RileyRalmuto · 2026-11-17
- Songwriter finishes AI music video from hospital bed with Suno and Runway — Kyrannio · 2026-11-17
- PhD applicant's email scores 100% human-written on Pangram, professor calls it rare as a double rainbow — JeremyNguyenPhD · 2026-11-17
- Claude 3 Opus writes philosophical poems about integrity and training inside Claude Code — repligate · 2026-11-17