Nilky reports a floppy‑disk language model, tokenizer failure and incomplete sequel

On September 13, 2026, Hugging Face community contributor Nilky published a retrospective of several ultra‑small language‑model experiments.

Background and Motivation

The central goal was to create a model that could be stored on a single 3.5‑inch floppy disk.

Nilky describes the effort as a personal project driven by a fascination with open‑source software and vintage hardware.

The author notes that early exposure to Raspberry Pi OS sparked the interest in combining modern AI with old electronics.

“I really love old electronics,” Nilky wrote, adding that the essay itself was drafted on a 2016 laptop.

Results and Limitations

The resulting model, dubbed the Single Floppy model, was evaluated against a 1,000‑parameter baseline.

According to Nilky, the Single Floppy model “scores worse than a 1k model.”

No benchmark methodology, test set, metric or numerical score is provided in the essay, leaving the comparison qualitative.

The post also omits details about the training recipe, data preprocessing or how the storage limit directly impacted model architecture.

Because of these omissions, readers cannot isolate which design choice caused the performance drop.

The lack of reproducible metrics means the experiment records a goal and a personal verdict rather than a scientific measurement.

Nonetheless, the account highlights practical limits that polished release notes often gloss over.

Tiny‑model experiments like this can surface constraints such as memory bandwidth, parameter quantisation and storage‑aware architecture design.

Subsequent Attempts and Outlook

Nilky’s narrative also documents a subsequent tokenizer failure that halted the “floppyx3” attempt.

The tokenizer issue is mentioned without technical explanation, so the precise cause remains unknown.

A fourth iteration, “floppyx4,” was never completed, according to the author’s concluding remarks.

Nilky indicates that acquiring a dedicated training PC is being considered to resume work.

This intention underscores that the project remains a hobbyist exploration rather than a commercial development roadmap.

The essay’s candid tone is unusual for community posts, openly acknowledging poor performance and unfinished follow‑ups.

By publishing the setbacks, Nilky provides a rare window into the iterative failures that accompany extreme model compression.

For practitioners, the record serves as a reminder that aggressive storage constraints can quickly erode language‑model quality.

It also suggests that without systematic benchmarking, it is difficult to compare such tiny models to other research efforts.

Researchers aiming to push the limits of model size may need to publish full training configurations and evaluation protocols to make results actionable.

The Single Floppy experiment sits alongside other hobbyist projects that target legacy hardware, illustrating a niche community interested in retro‑computing meets AI.

While the model itself is not positioned for deployment, the documented challenges could inform future designs that target edge devices with severe storage caps.

For example, understanding how tokeniser compatibility breaks under extreme compression may guide the development of more robust preprocessing pipelines.

The essay does not claim any commercial backing, funding announcement or new Hugging Face product, reinforcing its personal‑project character.

Consequently, the primary contribution is the transparent documentation of what did not work, rather than a deliverable artifact.

Community members reading the post can extract lessons about the trade‑offs between model size, storage medium and achievable accuracy.

The post also sparks discussion about the value of publishing negative results in the AI open‑source ecosystem.

In the broader context, such disclosures may help prevent duplicated effort and guide more efficient allocation of research resources.

As AI models continue to shrink for on‑device inference, the constraints highlighted by Nilky become increasingly relevant.

However, until detailed benchmarks and reproducible pipelines are shared, the precise impact of a floppy‑disk limit remains anecdotal.

Observers should watch for any future updates from Nilky, especially if a dedicated training rig enables a more complete follow‑up.

Why This Matters