← Back to blog
AiAbout 7 min read

Someone rebuilt IDM from the outside and wrote down what they did not touch

Published Nov 15, 2026
Someone rebuilt IDM from the outside and wrote down what they did not touch

A very old window with new parts inside

Download managers occupy a strange corner of Windows. A browser can already fetch a file; the manager exists to open six connections to the same server instead of one, reassemble the pieces, and resume when the wifi drops. Internet Download Manager has done exactly this since the early 2000s and still asks for money after a 30 day trial.

Mister Download Manager is a Rust reimplementation of that workflow, and it is deliberately familiar. Toolbar, categories, download list, progress window, browser button. The README says the point is that there is "nothing new to learn", an odd selling point, and the right one here. The repo is anefzaoui/misterdownloadmanager, GPL-3.0, 10,062 stars as of 15 November 2026. Builds cover Windows 10 and 11 (x64, ARM64, 32-bit), Linux (Ubuntu, Debian, Mint, Fedora, openSUSE, Arch and others, on x86-64 and ARM64), and macOS for Apple silicon and Intel, marked experimental. 86 languages. No ads, no trial, no sign-up.

The 5x, and the row underneath it

The README's table was measured, it says, on one computer, same files, both programs at default settings:

Mister Download ManagerIDM 6.42
1 GB file1 second5.5 seconds
Small fileinstantly3.1 seconds
Shaky connection1.6 seconds5.4 seconds
Wait before the download startsinstantly3.1 seconds
Pricefree, foreverpaid after a 30 day trial
Open sourceyes, GPLno

I do not believe the first row, and you should not either. One gigabyte in one second is about 8 Gbps of aggregate throughput. Nobody has that at home. The benchmark ran against the project's own labserver crate, a local HTTP test server with fault injection, so the absolute numbers describe a lab, not a broadband line. They measure how two engines behave when the server is nowhere near the bottleneck. That is a fair test of a download engine and a useless test of what you will see on a 200 Mbps cable connection.

The row that carries information is the shaky connection: 1.6 seconds against 5.4, same file, same server, simulated network trouble. If the engine really handles packet loss and stalled sockets better, that advantage survives the trip to real networks.

Then there is the detail the table leaves out. On a server that allows only two connections, MDM used two and still finished first. IDM opened 480. Plenty of download accelerators get their speed by being rude, hammering a server with as many sockets as it will tolerate until something gives. A program that opens 480 connections against a limit of two is not accelerating, it is retrying and hoping. MDM claiming a win at two connections is claiming the segment scheduler is better at deciding which piece to fetch next, and that claim is falsifiable in a way that "1 second" is not.

Two horizontal glass tubes of different bore, the same amount of dark sand already piled at the end of the narrow one.

The harness lives in tools/bench and the raw data in docs/bench/run5/REPORT.md. That is the only reason to take any of this half seriously. A benchmark run by the author, on the author's machine, against a competitor's default settings, is a marketing artifact with a chart. Publishing the harness turns it into something a stranger can break, and someone eventually will.

What clean-room means when it stops being a legal department

The term comes from semiconductor manufacturing, where the room has no dust in it. In software the room is kept free of the original code. Two teams work in isolation: one may read the competitor's product and write a specification, the other may write code but never look at the thing being cloned. Compaq used the arrangement in 1982 to produce a BIOS compatible with IBM's PC without copying it. Phoenix did the same work and sold the result to everyone else. ReactOS has been applying it to Windows for a very long time. The method turns up wherever a company wants compatibility without a lawsuit, and it always arrives with lawyers, memos and a paper trail long enough to prove the two teams never met.

Which is what makes MDM strange. The paper trail here is a bulleted list in a public README, written by one person, in the tone of someone explaining himself to a stranger.

The paragraph, read closely

In its own terms: there is no code, binaries, artwork, icons, sounds or text from Internet Download Manager in the project, and no code from any other download manager. The app icon was made for MDM with ChatGPT; the other icons come from openly licensed sets. No IDM file was ever disassembled, decompiled, debugged, patched or inspected below its user interface, and its licensing, trial and registration were never touched or bypassed. What MDM does was learned from IDM's public documentation and help pages, from public screenshots, and by running a licensed copy as a black box against the project's own test server. The download engine follows RFC 9110 and RFC 9112 and was designed, tested and measured independently. MDM is not affiliated with Tonec Inc.

The black-box bullet is the load-bearing one. Watching a licensed copy is what keeps a project on the right side of the line; opening the executable puts it on the wrong side. The RFC bullet does similar work, because it relocates the defence. The claim is not "we did not copy the code", which is hard to prove and easy to doubt. The claim is "we did not need the code, the behaviour is specified in a public standard and we implemented the standard". That is the stronger position, and the reason the 5x claim and the legal paragraph are the same argument. If the speed comes from HTTP range requests and a scheduler, it can be rebuilt from a document. If it came from somebody's binary, it could not.

I cannot verify any of this from outside. What I can say is that the paragraph is specific in the places where vagueness would be damning. It names the icon source, the RFC numbers and the trademark holder, and it volunteers that a licensed copy of a paid program was used during development, which is the admission a weaker document would leave out.

What you get, and what is missing

The feature list runs long in a way that suggests the author kept fixing his own irritations. Browser extensions for Firefox and Chrome (the Chrome Web Store listing also serves Edge, Brave, Opera, Vivaldi and Helium). Video downloads through yt-dlp, audio only included. Image galleries through gallery-dl. Torrents and magnet links through librqbit, with file selection, seeding, and tabs for peers, trackers and pieces. IDM's command line switches, plus its own CLI and a local API other programs can call. It is a Cargo workspace on edition 2024.

A download manager that also handles magnet links removes a second install, and the second install usually has the worse interface.

The 86-language count matters. IDM ships 40. For a utility, localisation is not decoration: a download manager is the kind of program people install for a parent, and a menu in the wrong language turns a five minute job into a phone call. The README is honest that most of the 86 are machine-made, marked #, machine in the translation files and waiting for native speakers, with Malay's menus reviewed by one. Eighty-six languages that are mostly unreviewed is a different product from eighty-six that are good. Machine translation of "Pause" and "Resume" still beats an English menu.

Against that: no code signing yet. The README warns that Windows will show "Windows protected your PC" on first run and tells you to click through. For anyone trained to stop at that dialog, this is a bigger barrier than any benchmark number.

Where I would start

The useful thing in this repository sits in docs/bench/run5/REPORT.md, next to a harness anyone can point at a server they control. Two connections on a throttled server, repeated on hardware that is not the author's, would settle more in an afternoon than the table settles in a year. Until then, the 5x belongs in the same drawer as every other vendor benchmark, and the clean-room section is the part worth reading twice.

Related articles