:: commit b487fe18225728c05442608853c28e7fadfc15b5

diff --git a/doc/file_format.md b/doc/file_format.md
index a7dcdf2..68c354f 100644
--- a/doc/file_format.md
+++ b/doc/file_format.md
@@ -22,6 +22,11 @@ static void write_neutral_s32(u8 * data, s32 value) {
 
 ## The chunks
 
-After the header, the file may contain an unlimited amount of chunks. Each chunk starts with the _new_ size - a 32-bit integer signifying the _compressed_ size of the block, and the _old_ size - a 32-bit integer signifying the _decompressed_ size. Then, a sequence of bzip3-compressed data follows.
+After the header, the file may contain an unlimited amount of chunks. Each chunk starts with the _new_ size - a 32-bit integer signifying the _compressed_ size of the block, and the _old_ size - a 32-bit integer signifying the _decompressed_ size. Then, a sequence of bzip3-compressed data follows. CRC32 checking is left up to libbz3.
 
-CRC32 checking is left up to libbz3.
+If the chunk is smaller than 64 bytes, then compression is not attempted. Instead, the content is prepended with the 32-bit CRC32 checksum and a 0xFFFFFFFF literal.
+
+Otherwise, the chunk starts with the 32-bit CRC32 checksum value, the Burrows-Wheeler transform permutation index and the compression _model_ - a 8-bit value specifying the compression preset used. As such:
+
+- 2-s bit set in the _model_ - LZP was used and the 32-bit size is appended to the block.
+- 4-s bit set in the _model_ - RLE was used and the 32-bit size is appended to the block.