QR code capacity and error correction, without the guesswork
Capacity is fixed by two choices: the symbol version, which runs from 1 to 40, and the error-correction level, L to H. In byte mode a version 1 symbol at level L holds 17 bytes; a version 40 symbol at level L holds 2953. Raising the level to H cuts version 1 down to 7 bytes for the same symbol size, because a third of the codewords now carry recovery data instead of your content.
Those numbers come from the standard, not from a blog post — and they were re-measured for this page by running each combination through the encoder the generator uses.
Why the two settings fight each other
Every symbol is a grid of modules. Version 1 is 21×21 and each version adds four modules per side, so version 40 is 177×177. Function patterns — three finder squares, timing lines, alignment patterns, format and version information — eat a fixed slice of that grid. Whatever is left becomes codewords, eight bits each.
Error correction does not change the number of codewords. It changes how they are spent. At level L, about 7% of the codewords are Reed–Solomon parity; at H, about 30%. The data capacity falls accordingly, and the parity codewords are what let a scanner reconstruct the message when part of the symbol is dirty, torn, curved or covered by glare.
So "more error correction" is not free robustness — it costs capacity, and if the payload no longer fits you are pushed to a higher version, which means more modules in the same physical space and a harder job for the camera.
The capacity table
Maximum characters in byte mode, which is what any UTF-8 text, URL or WiFi payload uses:
| Version | Modules | L (7%) | M (15%) | Q (25%) | H (30%) |
|---|---|---|---|---|---|
| 1 | 21×21 | 17 | 14 | 11 | 7 |
| 2 | 25×25 | 32 | 26 | 20 | 14 |
| 3 | 29×29 | 53 | 42 | 32 | 24 |
| 5 | 37×37 | 106 | 84 | 60 | 44 |
| 7 | 45×45 | 154 | 122 | 86 | 64 |
| 10 | 57×57 | 271 | 213 | 151 | 119 |
| 20 | 97×97 | 858 | 666 | 482 | 382 |
| 30 | 137×137 | 1732 | 1370 | 982 | 742 |
| 40 | 177×177 | 2953 | 2331 | 1663 | 1273 |
Read it as a ceiling, not a target. A version 40 code is a 177×177 grid; printed at business-card size each module is a fraction of a millimetre and no phone will resolve it. The practical ceiling for anything a person scans at arm's length is around version 10 to 15.
Numeric and alphanumeric modes hold more
Byte mode is the least efficient of the four encoding modes, and it is the one your text almost always uses. The standard defines three others:
- Numeric — digits only, packed three at a time into ten bits. A version 40-L symbol holds 7089 digits.
- Alphanumeric — digits, uppercase A–Z, space and
$ % * + - . / :, two characters per eleven bits. Version 40-L holds 4296 characters. - Kanji — Shift JIS double-byte characters, version 40-L holds 1817.
The catch: a mode only applies if the entire payload fits its character set. One lowercase letter or one ? drops the whole string back to byte mode, because a URL like https://example.com/qr/a is lowercase and therefore not alphanumeric. In practice, byte mode is what you get, and the byte table above is the one that matters.
Why non-Latin text runs out faster
Byte mode counts bytes, not characters. In UTF-8, an accented Latin letter is 2 bytes, most CJK characters are 3, and an emoji is 4. A contact card with a Chinese name and a Japanese company name burns through capacity roughly three times faster than the same card in English.
This is also where a lot of encoders silently corrupt data: if the library maps characters to single bytes instead of UTF-8, the symbol decodes to mojibake. The generator here switches the encoder to its UTF-8 path for exactly this reason, and the payload box under the preview shows what will actually be encoded.
Choosing an error-correction level
| Level | Recovers | Use it for |
|---|---|---|
| L | ~7% | Clean digital displays, high-resolution screens, controlled lighting |
| M | ~15% | The default for print: business cards, menus, flyers |
| Q | ~25% | Packaging, stickers, anything handled or exposed to weather |
| H | ~30% | Harsh environments, curved surfaces, codes with a logo knocked out of the middle |
M is the right default because it buys meaningful damage tolerance without inflating the symbol. Jump to H only when you know why: a code printed on a coffee cup lid, a label that will be scuffed in a warehouse, or an artwork with a logo covering the centre.
One caveat worth stating plainly: error correction repairs missing modules, not bad decisions. A low-contrast code at level H still fails, because there is nothing to repair — the camera cannot see the modules in the first place.
Fitting more into the same symbol
If your payload is too long, four changes usually solve it before any technical trick does:
- Strip tracking parameters. Campaign URLs with five UTM fields can be longer than the page address itself.
- Drop the scheme where it is not required. Scanners insert
https://when they see a bare domain, though this is less consistent than encoding it. - Use your own short path if you control the domain —
example.com/mbeats a long query string, and unlike a third-party shortener it does not expire. - Split the content. Two codes side by side — one for the menu, one for the WiFi — keep both symbols small and readable instead of forcing one dense grid.
If none of that is enough, the payload is asking too much of the medium. A 2900-byte text document in a QR code is technically possible and practically unscannable.
Frequently asked questions
What is the absolute maximum a QR code can hold?
Version 40 at level L: 7089 numeric digits, 4296 alphanumeric characters, 2953 bytes in byte mode, or 1817 Kanji characters. Every one of those maxima requires a 177×177 symbol, which needs a large, high-resolution print to be readable.
Does a denser code take longer to scan?
Marginally, on modern phones. The bigger effect is reliability: more modules in the same area means smaller modules, and small modules are the first thing to disappear under motion blur, low light or a scratched print.
Does a higher error-correction level make the code physically bigger?
At a fixed version, no — the symbol is the same grid with fewer data codewords. At a fixed amount of content, yes: the extra parity pushes you to a higher version, so the symbol gains modules.
Generate a code and watch the version, module count and character count update live as you type — the fastest way to feel where the capacity wall is. For the physical side of the same problem, read print size and quiet zone.
Written by the QRQuick team. Last updated November 19, 2026.