18 - Nov - 2026

Windows has a faster way to copy thousands of tiny files and File Explorer isn’t it

I have noticed a certain File Explorer behavior when I copy project files. The progress bar sprints to around 90%, after which it stalls and crawls until the job is completed. I don’t have the same experience copying single giant video files, and it has always felt backward.

I simply assume it’s Windows just being Windows. However, I tested Robocopy, a tool built into Windows, against Explorer, and it was faster, though not exactly in the way I expected.

300MB took longer than 5GB

Separate files played a significant role

File-explorer-small-files-test
Afam Onyimadu / MUO

On my computer, it took File Explorer about 30 seconds to copy about 5,000 text files, roughly 300MB in total. That works out to an average of about 10MB/s. However, on that same drive, I was able to copy a single 5GB file at around 260MB/s.

This is an experience similar to moving 5,000 envelopes compared with one heavy box. What may make the former slower is the paperwork per-item overhead. On a computer, the individual files require their own open, write, close, and bookkeeping steps. When the files are this small, the work involved in handling each individual file can become a significant part of the total copy time. Explorer spent about three seconds “calculating” before it started copying.

The setup I ran on Explorer included 5,000 generated text files on a WD SN530 NVMe SSD, running Windows 11 25H2 (build 26200.9457). I pointed Robocopy at the same folder. Windows already includes this tool.

Robocopy only beat File Explorer once I switched on one setting

Windows ships with Robocopy, a command-line copier, so you are not installing anything to use it. You simply open a command prompt and run Robocopy. When I forced Robocopy to use a single thread with /MT:1, there was barely a difference. For the same files, it took 29 seconds compared to Explorer’s 30 seconds.

However, multithreading changes everything. Here is the exact command I used:

robocopy "C:Source" "D:Destination" /E /MT:16 /NFL /NDL /NP /LOG:C:copy.log

/E tells it to copy subfolders, while /MT:16 allows 16 files to be copied concurrently. Together, the last four—/NFL, /NDL, /NP, and /LOG:C:copy.log—disable live progress output, reducing console rendering overhead. However, I had to open the log if I needed to see what happened.

While testing, I was tripped up by two things. The source has to come before the destination. When I deliberately swapped them, Robocopy treated the second path as the destination and copied in the opposite direction. It didn’t stop to ask whether I’d made a mistake. It also copies the contents of the source folder into the destination, rather than creating a Source folder inside it.

My command used 16 threads, and that same copy took about a quarter less time — down from 30 seconds to 22 seconds. Your day doesn’t suddenly change because you saved eight seconds; however, it made me curious. If 16 threads handed me eight seconds, I needed to know what 128 threads would do.

128 threads performed similarly to 4 threads

Speed stopped improving past 16, and one huge file didn’t improve at all

robocopy-mirror-preview-2
Afam Onyimadu / MUO

I didn’t quite get the results I was hoping for. When I updated my command to /MT:128, the operation lasted 25 seconds, similar to what I got when I used /MT:4.

Setting

Time (Seconds)

Time saved vs. Explorer

Explorer

30

baseline

/MT:1

29

~3%

/MT:4

25

~17%

/MT:8

24

20%

/MT:16

22

~27%

/MT:32

22

~27%

/MT:64

23

~23%

/MT:128

25

~17%

/MT:1 (rerun last)

29

~3%

The table above summarizes my results for ~5,000 generated text files (~300MB). The tests were done on the same NVMe SSD, and the destination was cleared before each run. I recorded the median of three runs.

I saw steady gains up to 16 threads, after which it flattened. Since one second lands within my run-to-run noise, I wouldn’t rank 16, 32, and 64 against each other. I needed to be sure that a warm file cache wasn’t inflating the latter results, so I reran Explorer and /MT:1. Once again, the results were 30 and 29 seconds.

I can’t definitively ascertain the cause of the plateau. My best guess is that the drive, antivirus scanning, or another per-file bottleneck was becoming the limiting factor.

There was no performance gain on the single 5GB file. On Explorer, it took about 19 seconds. It took an equal amount of time with /MT:1, /MT:16, and /J too, so none of those switches improved this particular transfer. A single file doesn’t give /MTmultiple files to work on concurrently, so there was no similar opportunity for the extra threads to help.

Used correctly, and for the right kinds of files, I’ll stick with Robocopy, but there is one thing to bear in mind.

One setting that slows it down and a command that can erase files

This entire experiment was around a single NVMe drive and generated text files. I wouldn’t use it to predict how photos, hard drives, USB sticks, or network shares behave.

I tried a few other switches:

Switch

What it does

What happened

The catch

/Z

Restartable mode, survives interruptions

~25 s vs ~22 s

~14% slower; worth considering when a copy might be interrupted

/R:1 /W:1

One retry, one-second wait

No locked or failed files

Default is 1,000,000 tries at 30s (in theory, a very long stall)

/MIR

Mirrors source, deletes extras at destination

Disposable data only

Can erase files; use /L first

You should pay attention to /MIR. It deletes anything extra as a way of making the destination match the source. If you swap paths, you may end up wiping the wrong folder.

The default retry count is 1,000,000 (/R) with a 30‑second wait (/W) between attempts, which explains why a specific copy can seemingly hang forever.

I lean more towards using Robocopy for several small-file copies, but the useful part was learning that more threads eventually stopped helping, while a few switches could make a copy slower or even delete files if I wasn’t careful.

Leave a Reply

Your email address will not be published. Required fields are marked *