Usenet client download benchmarks

The raw numbers from a set of client download rounds, without commentary. Each table races one current build of each client on identical hardware, the same five providers, TLS everywhere, runs interleaved, and every leg's output checked byte for byte. Rows are listed in a fixed order in every table, not ranked by result. Nothing on this page is highlighted.

Method

How the numbers were taken

Builds: NZBGet 26.3-testing, SABnzbd 5.1.1, nzbfast 1.2.2, Weaver 0.7.8 (its newest shipped binary, identity checked by hash), rustnzb 1.4.5. One build of each client across every table on this page.
Configuration: each client at its documented best settings. SABnzbd ran pipelining_requests=8; NZBGet ran ArticleCache, DirectWrite, DirectUnpack and ParQuick; rustnzb ran its documented config. nzbfast ran as shipped: on the 1 Gbit and slower rounds its connection governor held 25 connections; on the 10 GbE round it ran 50 connections (10 per server). The other clients ran their configured connection counts, in the hundreds.
End state: the clock stops at a verified, extracted, usable file, not at "download finished".
Measurement: elapsed time, process-tree peak RSS and CPU time sampled at 2 Hz, device I/O in GiB, wire bytes in decimal GB. Three repetitions per client per fixture with the order rotated; the figure quoted is the median unless a footnote says otherwise. Every leg's output was compared byte for byte against a reference checksum.
Hardware: the 250 Mbit, 500 Mbit and 1 Gbit rounds ran on a 20-core Apple Silicon machine (the two slower rates shaped from the 1 Gbit line, the shaped rate verified by an independent probe: 248 and 496 Mbit). The 10 GbE round ran on an Apple Silicon machine on a 10 GbE line. The Windows round ran on a native Windows 10 GbE machine with a TLC system drive sustaining 0.99 GB/s of writes.
Dates: the 1 Gbit round ran 23 August 2026; the 250 Mbit, 500 Mbit, 10 GbE, Windows and damaged-post rounds ran 24 August 2026.

250 Mbit line · 24 August 2026

6.5 GB release, 250 Mbit

client	time to usable file	peak memory (RSS)	CPU time	device I/O (GiB)
NZBGet 26.3-testing	227 s	826 MB	78.4 s	13.3
SABnzbd 5.1.1	230 s	1,667 MB	162.4 s	15.2
nzbfast 1.2.2	217 s	144 MB	53.3 s	6.2
Weaver 0.7.8	230 s	789 MB	62.0 s	12.9
rustnzb 1.4.5	256 s	644 MB	122.4 s	15.4

Medians of three, order rotated, every leg byte-checked. Byte counts on a shaped line are read from each client's own counter, never the network interface (the shaper drops and TCP resends, so the interface counts both copies); no wire column is quoted for the shaped rounds for that reason.

500 Mbit line · 24 August 2026

6.5 GB release, 500 Mbit

client	time to usable file	peak memory (RSS)	CPU time	device I/O (GiB)
NZBGet 26.3-testing	145 s¹	846 MB	64.3 s	13.2
SABnzbd 5.1.1	125 s	1,665 MB	99.2 s	14.2
nzbfast 1.2.2	109 s	142 MB	45.4 s	6.2
Weaver 0.7.8	138 s	564 MB	48.5 s	12.4
rustnzb 1.4.5	131 s	626 MB	92.6 s	14.5

Medians of three, order rotated, every leg byte-checked. The nzbfast arm here is the v1.2.2 release tag build. ¹ NZBGet's three legs spanned 114 to 158 s with flat resource readings; the median is quoted and the spread stated.

1 Gbit line · 23 August 2026

6.5 GB named release, 1 Gbit

client	time to usable file	peak memory (RSS)	CPU time	device I/O (GiB)	wire (GB)
NZBGet 26.3-testing	61 s	801 MB	40.1 s	12.5	6.5
SABnzbd 5.1.1	63 s	1,588 MB	69.5 s	13.8	6.5
nzbfast 1.2.2	58 s	143 MB	35.7 s	6.2	6.5
Weaver 0.7.8	110 s	531 MB	34.9 s¹	12.5	6.5
rustnzb 1.4.5	67 s	628 MB	61.9 s	13.4	7.2

34 GB obfuscated release, 1 Gbit

client	time to usable file	peak memory (RSS)	CPU time	device I/O (GiB)	wire (GB)
NZBGet 26.3-testing	306 s	900 MB	221.3 s	68.1	34.4
SABnzbd 5.1.1	308 s	1,607 MB	356.5 s	72.5	34.4
nzbfast 1.2.2	302 s	191 MB	194.0 s	32.5	34.4
Weaver 0.7.8	511 s	1,076 MB	373.8 s	98.0	34.4
rustnzb 1.4.5	343 s	445 MB	332.6 s	69.8	37.8

Medians of three, order rotated, 30 of 30 legs on these two fixtures byte-identical. The nzbfast arms ran the build that shipped later the same night as v1.2.2, at shipped defaults (25 connections). ¹ Weaver's three CPU legs on the 6.5 GB fixture spanned 34.1 to 56.1 s; the median is quoted and the spread stated. rustnzb's 7.2 and 37.8 GB wire figures are its own counter's excess over the 6.5 and 34.4 GB posts, with a warning in its log on two legs.

10 GbE line · 24 August 2026

87 GB post to a usable 77 GB file, 10 GbE

client	time to usable file	peak memory (RSS)	CPU time	device I/O (GiB)	wire (GB)
NZBGet 26.3-testing	93 s	1,195 MB	443 s	183.7	77.2
SABnzbd 5.1.1	113 s	1,910 MB	234 s	235.1	77.2
nzbfast 1.2.2	70 s	415 MB	144 s	72.9	77.2
Weaver 0.7.8	645 s	1,825 MB	631 s	435.1¹	77.3
rustnzb 1.4.5	869 s²	681 MB	2,444 s	216.1	86.9

Medians of three, order rotated, 15 of 15 legs on this fixture byte-identical across these five arms, all clients agreeing on the output checksum. The nzbfast arm is the v1.2.2 release build at 50 connections (10 per server); a six-rung sweep on this fixture measured the wall identical from 50 connections to 360. Rep 3 for the three fastest arms ran about 40 minutes after the rest (a free-space guard paused the round; the offset moved no median by more than the rep spread). ¹ Weaver's device I/O is its encrypted-at-rest store re-reading and re-writing data as the job grows. ² rustnzb completed byte-correct on all three legs; it pulled 86.9 GB where the eager plan is 77.2 (it fetches the full recovery set unconditionally).

Windows · 24 August 2026

The same 87 GB post on native Windows

client	time to usable file	peak memory (RSS)	CPU time	device I/O (GiB)	wire (GB)
Newsbin Pro 6.90 (360 connections)	477 s	881 MB	1,862 s	154.1	76.7
nzbfast 1.2.2 (50 connections)	105 s	469 MB	154 s	84.5	76.7

Medians of three, interleaved, 6 of 6 legs byte-checked against the same checksum as the table above. Native Windows 10 GbE machine, TLC system drive sustaining 0.99 GB/s of writes. Newsbin ran at its own configured per-server connection maxima, and its times exclude the 90-second settle the harness waits before declaring a watched client finished. Newsbin is a registered trademark of CMCE, Inc.; the client is published by DJI Interprises, LLC.

Damaged post · 24 August 2026

6.5 GB release with 60, 20 and 5 dead articles

time to a verified, usable file (mean of 3)	60 dead articles	20 dead	5 dead
NZBGet 26.3-testing	33.0 s	23.7 s	23.0 s
SABnzbd 5.1.1	48.0 s¹	28.0 s	24.3 s
nzbfast 1.2.2	13.0 s	10.0 s	8.7 s
Weaver 0.7.8	did not finish²	did not finish²	194.0 s
rustnzb 1.4.5	107.3 s	51.3 s	47.0 s

10 GbE machine, five providers, 100 connections per client, the same 6.5 GB release poisoned at three damage levels, three repetitions per arm with the order alternated, every leg byte-checked against the clean file. Means of three. ¹ SABnzbd's three legs at the heaviest damage ran 43, 41 and 60 s; the mean is quoted and the range stated. ² Weaver moved 1.5 GB and 4.9 GB of the 6.5 within the 20-minute cutoff at the two heavier damage levels, and completed correctly all three times at 5 dead articles.