QR error correction capacity table: data vs EC codewords

Every QR version has a fixed codeword budget split between data and error correction. Version 1 has 26 codewords, 19 data and 7 error correction at level L, but 9 and 17 at level H. Version 40 ranges from 2,956 data codewords at L to 1,276 at H. The table lists all forty versions.

How to read the table

A codeword is 8 bits. Each version has a fixed total (version 1 always carries 26 codewords, version 40 always 3,706), and the error-correction level decides how that total is split between data and Reed–Solomon redundancy. Cells below are data / EC.

These values were extracted from the RS block table of the encoder UseQR itself uses, which implements the error-correction characteristics table of ISO/IEC 18004. They are pulled from running code, not transcribed from a blog.

All forty versions

Version	Total	L (data / EC)	M (data / EC)	Q (data / EC)	H (data / EC)
1	26	19 / 7	16 / 10	13 / 13	9 / 17
2	44	34 / 10	28 / 16	22 / 22	16 / 28
3	70	55 / 15	44 / 26	34 / 36	26 / 44
4	100	80 / 20	64 / 36	48 / 52	36 / 64
5	134	108 / 26	86 / 48	62 / 72	46 / 88
6	172	136 / 36	108 / 64	76 / 96	60 / 112
7	196	156 / 40	124 / 72	88 / 108	66 / 130
8	242	194 / 48	154 / 88	110 / 132	86 / 156
9	292	232 / 60	182 / 110	132 / 160	100 / 192
10	346	274 / 72	216 / 130	154 / 192	122 / 224
11	404	324 / 80	254 / 150	180 / 224	140 / 264
12	466	370 / 96	290 / 176	206 / 260	158 / 308
13	532	428 / 104	334 / 198	244 / 288	180 / 352
14	581	461 / 120	365 / 216	261 / 320	197 / 384
15	655	523 / 132	415 / 240	295 / 360	223 / 432
16	733	589 / 144	453 / 280	325 / 408	253 / 480
17	815	647 / 168	507 / 308	367 / 448	283 / 532
18	901	721 / 180	563 / 338	397 / 504	313 / 588
19	991	795 / 196	627 / 364	445 / 546	341 / 650
20	1085	861 / 224	669 / 416	485 / 600	385 / 700
21	1156	932 / 224	714 / 442	512 / 644	406 / 750
22	1258	1006 / 252	782 / 476	568 / 690	442 / 816
23	1364	1094 / 270	860 / 504	614 / 750	464 / 900
24	1474	1174 / 300	914 / 560	664 / 810	514 / 960
25	1588	1276 / 312	1000 / 588	718 / 870	538 / 1050
26	1706	1370 / 336	1062 / 644	754 / 952	596 / 1110
27	1828	1468 / 360	1128 / 700	808 / 1020	628 / 1200
28	1921	1531 / 390	1193 / 728	871 / 1050	661 / 1260
29	2051	1631 / 420	1267 / 784	911 / 1140	701 / 1350
30	2185	1735 / 450	1373 / 812	985 / 1200	745 / 1440
31	2323	1843 / 480	1455 / 868	1033 / 1290	793 / 1530
32	2465	1955 / 510	1541 / 924	1115 / 1350	845 / 1620
33	2611	2071 / 540	1631 / 980	1171 / 1440	901 / 1710
34	2761	2191 / 570	1725 / 1036	1231 / 1530	961 / 1800
35	2876	2306 / 570	1812 / 1064	1286 / 1590	986 / 1890
36	3034	2434 / 600	1914 / 1120	1354 / 1680	1054 / 1980
37	3196	2566 / 630	1992 / 1204	1426 / 1770	1096 / 2100
38	3362	2702 / 660	2102 / 1260	1502 / 1860	1142 / 2220
39	3532	2812 / 720	2216 / 1316	1582 / 1950	1222 / 2310
40	3706	2956 / 750	2334 / 1372	1666 / 2040	1276 / 2430

What each level recovers

Level	Nominal recovery	EC share of the symbol
L	~7%	19–27%
M	~15%	36–39%
Q	~25%	50–56%
H	~30%	63–66%

The EC share is roughly twice the recovery rate because Reed–Solomon spends two EC codewords to fix each codeword whose location is unknown, and one for an erasure whose location is known. At level H nearly two-thirds of the printed pattern is redundancy.

The smallest versions recover slightly less than the arithmetic suggests: version 1 at level L has 7 EC codewords but corrects only 2 codeword errors, because 3 of those 7 are reserved for misdecode protection rather than correction.

Blocks, not one big calculation

From version 3 upward the data is split into multiple Reed–Solomon blocks that are error corrected independently and then interleaved. Version 40 at level H uses 81 blocks of 45–46 codewords each; version 1 is always a single block. Damage concentrated in one block can defeat a code even when the total damage is under the nominal percentage.

Turning codewords into characters

Multiply data codewords by 8 bits, subtract the 4-bit mode indicator and the character count indicator, and divide by the bits per character for the mode. That derivation is already done for you in the capacity table: version 40-L's 2,956 data codewords are exactly where the famous 7,089-digit ceiling comes from. To check a real payload rather than the theory, decode the rendered output.

FAQ

How many error correction codewords does a QR code have?

It depends on version and level: from 7 (version 1, level L) to 2,430 (version 40, level H). The total codeword count is fixed per version; the level only changes the data/EC split.

What percentage of a QR code is error correction?

Roughly 19–27% of codewords at level L, 36–39% at M, 50–56% at Q and 63–66% at H. That is about double the recovery rate, since fixing one unknown error costs two EC codewords.

Why does level H only recover 30% when 65% of the code is redundancy?

Reed–Solomon needs two redundant codewords per corrupted codeword when it does not know which ones are wrong. Only known-location erasures are corrected one-for-one.

Are these numbers from the QR specification?

Yes. They match the error-correction characteristics table in ISO/IEC 18004, and were extracted from the RS block table of the working encoder UseQR uses rather than copied from secondary sources.