Newshot

Google engineer runs a 270M-parameter AI model on a printer using about 14MB of memory, at roughly 60 seconds per token

AI summary

Omar Sanseviero, a member of Google DeepMind's technical team, posted on X on November 17 that he successfully ran the 270-million-parameter Gemma 3 270M model on an office laser printer using only 14.1MB of memory, at about 60 seconds per output token. Using Google's Gemini 4 Carbon model, he compiled a small Transformer inference engine and a quantized model into a PostScript print job, relying on the printer's built-in PostScript interpreter to perform the computation without modifying the Ricoh printer's firmware or adding an external compute board. To fit the printer's limited memory, the model was loaded dynamically layer by layer, with peak usage of about 14.1MB; prefill of the prompt "Read the fine" took about 184 seconds, and decoding a single token took 0.1 seconds.

Translation(model translation, not the source's own text)

Running a 270M-parameter model in ~14MB of memory: Google engineer puts AI inside a printer, ~60 seconds per token

IT之家 November 20 news: Omar Sanseviero, a member of the technical team at Google DeepMind, posted on X on November 17 that, using only 14.1MB of memory, he successfully got an office laser printer to run the 270-million-parameter Gemma 3 270M model, consuming about 60 seconds per output token.

With the help of Google's Gemini 4 Carbon model, Sanseviero explored and demonstrated a special experiment that compiled a small Transformer inference engine together with a quantized model into a PostScript print job.

The experiment did not modify the Ricoh printer's firmware, nor did it add an external compute board; instead, it used the printer's built-in PostScript interpreter to perform the computation, letting a language originally meant to describe printed pages take on the job of running AI.

To make the most of the printer's limited memory resources, the experiment did not hold the entire model in memory at once, but instead adopted dynamic layer loading, alternately paging each layer of the model into and out of memory; the experiment's peak usage during runtime was about 14.1MB. IT之家 attaches the relevant screenshot below:

On performance output, the experiment's processing speed was about 60 seconds per token (not 60 tokens per second). The report says the entire computation happens inside the printer, with the printer's CPU executing the PostScript program and then handing the answer to the printer to print; model inference involves 18 Transformer network layers.

In the prefill stage, processing the input prompt "Read the fine" took about 184 seconds, while decoding a single generated token took 0.1 seconds, for a total of 184.1 seconds.

Related articles

  • A phone with 2GB of memory can also use AI smoothly: Google lowers the barrier to using the Android Gemini App
  • Google launches Stitch Loop: positioned as a "product evolver" that never clocks off, watching your product and fixing bugs for you
  • Google Pics upgrades to the Nano Banana 2.1 model: more detailed image generation, more accurate image editing
  • The best AI assistant for an app side hustle: Google launches a series of tools aimed at the most painful part, monetization
  • A key step toward AI taking over phones: Google Gemini automation app list revealed, currently 134 apps in total
  • Letting AI reply directly to the user UI interface: Google's open-source cross-platform protocol A2UI version 1.0 goes live, ByteDance has already adopted it

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗