Engineer Runs Language Model on Office Laser Printer

A demonstration used a 270-million-parameter Gemma 3 model on a printer's built-in PostScript interpreter, with no firmware changes or added hardware.

Updated on Nov. 19, 2026 in Artificial Intelligence

Compact gray office laser printer on a bare wooden desk with a single blank sheet emerging from its output slot in a quiet office.
An engineer ran a 270-million-parameter Gemma 3 language model on an office laser printer, which printed the model's three-word reply without any firmware changes or added hardware. AI Illustration. Upload story photo >

An engineer ran a quantized version of Google's Gemma 3 270M language model on an office laser printer, and the printer physically printed the model's three-word response, "Read the fine print." The setup used the Carbon inference engine and PostScript, the page description language built into modern laser printers.

Why it matters

The demonstration shows that memory-efficient techniques such as dynamic layer streaming can let very small AI models run on hardware never designed for them. It also highlights how much unused computing capacity exists in everyday devices.

The run peaked at 14.1MB of RAM within 244.6MB of allocated PostScript virtual memory and produced output at 61.3 tokens per second on the printer CPU. Prefill of the four-token prompt took 184 seconds, while each generated token took 0.1 seconds, and the source does not indicate how these figures would change with longer prompts or larger models.

The players

Gemma 3 270M

A 270-million-parameter language model from Google's Gemma 3 family, used here in a quantized form.

Carbon

A lightweight transformer inference engine used to run the model on the printer.

The details

The engineer used a single raw PostScript print file that ran 18 transformer layers inside the printer's PostScript Level 3 interpreter. The system used dynamic layer streaming to cycle layers in and out of memory, and it did not modify printer firmware or attach external compute boards.

Timeline

  1. The printer demonstration was reported in an article published November 19, 2026.

The Tech Race

The demonstration repurposes PostScript, the page description language that has been built into laser printers for decades, as a general-purpose execution environment for a neural network. It shows that the same low-memory streaming approaches now used to fit small models onto phones and laptops can also reach devices whose main job is printing documents.

The demonstration shows that some AI features may eventually run on hardware people already own, such as printers, rather than requiring new purchases. Readers should not expect this approach to appear in consumer products soon, since the source describes it as a proof of concept with slow prefill times.

The takeaway

Efficient memory techniques are making it possible to run AI models on increasingly unusual hardware. Watching how these experiments scale will show whether everyday devices can take on more AI work.

Further reading

For more on the models and tools behind this kind of experiment, see Artificial Intelligence coverage.

Source note: This article includes information reported by Wccftech.