Gemma 3 270M Runs Inside an Office Printer at 61.3s a Token

What Was Posted on X

Omar Sanseviero of Google DeepMind posted on 17 November that he ran Gemma 3 270M directly in an office printer, with 14MB peak RAM. The printout in his photo is the source for the timings and memory figures in this piece.

Google DeepMind’s Omar Sanseviero ran Gemma 3 270M entirely inside an office laser printer, he posted on 17 November. The printout he photographed shows 184.0 seconds of compute, or 61.3 seconds per token.

That is seconds per token, not tokens per second. Wccftech reported “61.3 tokens per second”; the page in the photo reads “61.3 s/token”.

Business Pill · GPUs Explained

An evergreen Business Engineer explainer on GPUs. It teaches the general idea only and does not discuss this printer demo.

The key insight: Gemma 3 270M ran inside a Ricoh office printer at 61.3 seconds per token, per the printout Sanseviero photographed, using 14.1MB of the 244.6MB heap, about 5.8% (our arithmetic).

What Sanseviero Ran

Sanseviero wrote that “Carbon compiled a small transformer engine + quantized Gemma 3 270m into a single PostScript job”. He did not say in the post what Carbon is.

The model ran “100% in the printer”, he said, with “14MB peak RAM” and “Smart layers juggling to avoid OOMing”, that is, to avoid running out of memory.

The printout names the printer as a “RICOH IM C3010 PSJ” running PostScript Level 3. It lists 18 transformer layers and a “Toner-Saver Mode”.

The printout in Sanseviero's photo puts prefill at 184.0 seconds and decoding at 0.1 seconds, for 61.3 seconds
The printout in Sanseviero’s photo puts prefill at 184.0 seconds and decoding at 0.1 seconds, for 61.3 seconds per token.

What the Printout Shows

The prompt was four tokens, “Read the fine”. The model generated one more, and the page prints the result: “Read the fine print”.

Prefill took 184.0 seconds and decoding took 0.1 seconds, per the page. It puts the total at 184.0 seconds.

Memory peaked at 14.1MB of the 244.6MB allocated to the printer’s PostScript heap, about 5.8% (our arithmetic). The page labels the method “Auto Low-RAM Layer-Streaming”.

Why a 270M Model Fits

Gemma 3 270M is the 270-million-parameter compact model Google introduced in August 2025, built for small devices and task-specific fine-tuning. We covered the launch here.

Sanseviero’s demo puts that model on a device built to print pages, not to run software, which is why the run is measured in minutes rather than milliseconds.

The Structural Read

As we read the printout, almost all of the 184 seconds went to reading the prompt, not to writing the answer.

As we read the demo, small open models now fit into hardware that was never built to run them, which says more about the model than about the printer.

Omar Sanseviero, 17 Nov 2026

“With Carbon’s help, I ran Gemma directly in an office printer”

Three Implications

FOR DEVELOPERS Sanseviero says the run used 14MB peak RAM and layer juggling to avoid running out of memory.

FOR READERS OF THE COVERAGE The printout reads 61.3 s/token; Wccftech reported 61.3 tokens per second.

FOR GOOGLE Gemma 3 270M is the compact model Google introduced in August 2025.

What Is Not Established

The post does not explain what Carbon is or how the model was quantized. The printout covers one four-token prompt and one generated token, so it gives no speed for longer replies. Sanseviero calls the post his own views; it is not a Google product announcement.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer map of how AI capital, compute and power loop together.

Read the Map of AI →

The Bottom Line

A Google DeepMind engineer ran Gemma 3 270M inside a Ricoh office printer, his printout shows, at 61.3 seconds per token and 14.1MB of peak memory.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The quotes and the 14MB figure are from Omar Sanseviero’s post on X of 17 November 2026. The printer model, timings, token counts and memory figures are our reading of the printout in the photo attached to that post. The 5.8% is our arithmetic.

Sources: Omar Sanseviero, post on X, 17 Nov 2026 (with printout photo) · Google Developers Blog, Introducing Gemma 3 270M · Wccftech, 19 Nov 2026

Scroll to Top
FourWeekMBA