Llama zero-width space remover
What the Llama zero-width space remover does
A zero-width space is a character that occupies a position in your text and renders nothing at all. This tool scans text copied out of a Llama-based interface for U+200B and the small family of related invisible characters that travel with it, and removes them, leaving every visible word exactly where it was.
It runs entirely in your browser. Paste, press the button, copy the result: nothing is uploaded, there is no account, and there is no limit on how much text you clean or how often you come back.
Why it happens
Why the characters vary by which Llama frontend you used
Llama is a model family rather than a single hosted product, so a reply can reach you through a self-hosted web UI, a desktop app like Ollama or LM Studio, a hosted playground, or a custom application built on the model's API. Each one wraps the raw text in its own markdown renderer, and each renderer makes its own choices about formatting a list, a code block or a heading, choices that can introduce a non-breaking space or a zero-width character as a side effect.
The result is that the exact mixture of invisible characters in a Llama reply depends more on which frontend rendered it than on anything the underlying model produced, which is why the same answer can paste cleanly from one interface and carry hidden characters from another.
How it works
How to remove zero-width spaces from Llama text in three steps
There is nothing to preview beforehand, since the problem is invisible until you run the tool.
- Step 1
- Copy the text from whichever Llama interface you're using, and paste it into the box.
- Step 2
- Press Remove hidden characters. It runs instantly, on your own device.
- Step 3
- Copy the result. Every visible word is exactly as it was.
Worth knowing
This clears more than just the zero-width space
The zero-width space, U+200B, is the character most people mean by the name, but it rarely travels alone. This tool also removes its close relatives: the zero-width joiner and non-joiner, the word joiner, the byte-order mark, the soft hyphen, and the family of non-breaking, narrow and hair spaces that render as an ordinary gap but are a different codepoint underneath.
A joiner sitting inside a real emoji sequence is left alone, since removing it would visually split the emoji apart, and the space-like characters are swapped for an ordinary space rather than deleted outright, so wrapping and word boundaries stay correct.
Why it matters
What these hidden characters actually break
Inside a heading or a URL slug, a hidden character splits the string for anything comparing it byte by byte: search, a regular expression, a string equality check. Two values that look identical on screen simply fail to match, because one carries a character the other does not.
In a self-hosted pipeline that diffs model output between runs or prompts, a hidden character introduced by the rendering layer rather than the model can register as a change when the actual wording is identical, which is a frustrating source of noise in an automated comparison.
Straight answer
Does a hidden character mean Llama watermarked the text?
No. Llama is open-weight, and there is no single controlled copy path through which anyone could plant a deliberate marker. These characters are overwhelmingly a side effect of whichever frontend rendered the text, not evidence of who wrote it or which model produced it.
Separately, there is real research into watermarking AI text by biasing which words a model is statistically more likely to choose, a pattern buried in the writing itself rather than in any single character. No tool that scans for invisible Unicode, this one included, can see or affect that kind of statistical watermark.
Who uses it
Who uses the Llama zero-width space remover, and when
Developers building an application on a self-hosted Llama model run it before comparing output between runs, since a hidden character introduced by the rendering layer can register as a false difference in a diff.
Anyone doing programmatic comparison, deduplication, data validation, database imports, runs Llama-sourced text through here first so two values that should be identical actually match.