Choose a version:
35% The original file has 24615 bytes (24.0k) and is available from the project website.
There you can find the official minified version, too, which brings down the size to 8549 bytes (8.3k, 35%).

After GZIP compression these minified files vary in size:
Boot
  3228 bytes (3.2k)
CDN
gzip -6 (default)
  2944 bytes (2.9k)
local copy
Baidu
  2941 bytes (2.9k)
CDN
gzip -9
  2937 bytes (2.9k)
local copy
cdnjs
  2926 bytes (2.9k)
CDN
cdnhttps
  2920 bytes (2.9k)
CDN
7zip -mx=9 -tgzip
  2849 bytes (2.8k)
local copy
libdeflate -12
  2844 bytes (2.8k)
local copy
pigz -11 -n
  2843 bytes (2.8k)
local copy
kzip -s0 -rn -b1
  2841 bytes (2.8k)
local copy
Zopfli
  2839 bytes (2.8k)
local copy

perma-link to the smallest file on my server:
http://minime.stephan-brumme.com/files/underscorejs/underscore-1.0.2.min.js

You will automatically get the smallest UnderscoreJS 1.0.2 file, ETag caching is available and
if your browser doesn't support GZIP decompression then the uncompressed version will be sent.

Currently best Zopfli settings

Save 81 bytes by using my UnderscoreJS 1.0.2 Zopfli version instead of the best available CDN (2.85% smaller than cdnhttps, 2839 vs. 2920 bytes):
You can use my super-compressed files for whatever purpose you like as long as you respect the library's original license agreement.
There are no restrictions from my side - but please avoid hot-linking if you run a high-traffic website.

These command-line settings yielded the best compression ratio so far (Linux version of zopfli-krzymod):
zopfli --i10000 --mb8 --mls2048 --bsr18 --lazy --ohh

(found February 2, 2016)
Description Value Parameter
iterations 10000  --i10000
maximum blocks 8  --mb8
maximum length score 2048  --mls2048
block splitting recursion 18  --bsr18
lazy matching in LZ77 yes  --lazy
optimized Huffman headers yes  --ohh
initial random W for iterations 1  --rw1
initial random Z for iterations 2  --rz2

Verify file integrity

After decompression, my uncompressed files are identical to the original ones:

MD5:
curl --silent --compressed https://raw.githubusercontent.com/jashkenas/underscore/1.0.2/underscore-min.js --location | md5sum
e99377fb126528933b97f950123e5e58  -
curl --silent --compressed http://minime.stephan-brumme.com/files/underscorejs/underscore-1.0.2.min.zopfli.js.gz | md5sum
e99377fb126528933b97f950123e5e58  -

SHA1:
curl --silent --compressed https://raw.githubusercontent.com/jashkenas/underscore/1.0.2/underscore-min.js --location | sha1sum
7361e8afa72bd0098e2520584786474c9c4a5064  -
curl --silent --compressed http://minime.stephan-brumme.com/files/underscorejs/underscore-1.0.2.min.zopfli.js.gz | sha1sum
7361e8afa72bd0098e2520584786474c9c4a5064  -

These CDNs send you the original file:
CDN Size (compressed) MD5 (uncompressed) Timestamp
Boot 3228 bytes e99377fb126528933b97f950123e5e58 (invalid)
cdnjs 2926 bytes e99377fb126528933b97f950123e5e58 (invalid)
cdnhttps 2920 bytes e99377fb126528933b97f950123e5e58 (invalid)

And some CDNs send you a different file:
CDN Size (compressed) MD5 (uncompressed) Comment / Diff Timestamp
Baidu 2941 bytes ae9188a7bb626d6470a7f831403ecfbc only whitespaces differ (invalid)

Note: only the MD5 hashes are shown to keep things simple.

Other Versions

Available UnderscoreJS versions at minime.stephan-brumme.com:

1.9.0,
1.8.3, 1.8.2, 1.8.1,
1.7.0,
1.6.0,
1.5.2, 1.5.0,
1.4.4, 1.4.3, 1.4.2, 1.4.1, 1.4.0,
1.3.3, 1.3.2, 1.3.1, 1.3.0,
1.2.4, 1.2.3, 1.2.2, 1.2.1, 1.2.0,
1.1.7, 1.1.5, 1.1.4, 1.1.3, 1.1.2, 1.1.1, 1.1.0,
1.0.4, 1.0.3, 1.0.2, 1.0.1, 1.0.0,
0.6.0,
0.5.7, 0.5.6, 0.5.5, 0.5.4, 0.5.3, 0.5.2, 0.5.1, 0.5.0,
0.4.7, 0.4.6, 0.4.5, 0.4.4, 0.4.3, 0.4.2, 0.4.1, 0.4.0,
0.3.3, 0.3.2, 0.3.1, 0.3.0,
0.2.0,
0.1.1, 0.1.0

The project site contains an overview how well these versions were compressed.
Other interesting projects are AngularJS, BackboneJS, Bootstrap, D3, Dojo, Ember, jQuery, Knockout, lodash, React, Socket.IO, ThreeJS and Vue.

Changelog

Best Zopfli parameters so far:
Size Improvement Parameters Found
2839 bytes -1 byte zopfli --i10000 --mls2048 --bsr18 --lazy --ohh February 2, 2016 @ 20:45
2840 bytes zopfli --i100 --mls16384 --bsr25 --lazy --ohh February 2, 2016 @ 19:07

If there are multiple parameter sets yielding the same compressed size, only the first one found is shown.

Most recent activity on June 16, 2016 @ 14:35.

Heatmaps

This Zopfli heatmap visualizes how compression changes when modifying the --bsr and --mls parameter.
Cell's contents is the best filesize achieved (in bytes, hover with mouse over cells to see number of iterations).

Good parameters are green, bad are red. The best and worst are bold as well.
The brightness of the blue background color indicates how many iterations were processed:
10,000.
bsr \ mls
2 4 8 16 32 64 128 256 512 1024 2048 4096 8192 16384 32768
bsr \ mls
2 4 8 16 32 64 128 256 512 1024 2048 4096 8192 16384 32768
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839
2840 2840 2840 2840 2840 2839 2839 2839 2839 2839 2839 2839 2839 2839 2839

Due to the Monte Carlo design of my search algorithm, not all parameters have reached the same number of iterations yet:
Iterations Min. Bytes Reduction Coverage
100 2840 bytes 100%
1,000 2840 bytes 100%
10,000 2839 bytes -1 byte 100%
100,000
1,000,000
10,000,000

KZIP has far less options available for tuning/optimization. I only played around with the number of blocks (parameter -n):
Blocks Min. Bytes Compared To Best Zopfli Compared To Best KZIP
2841 bytes +2 bytes (+0.07%)
2841 bytes +2 bytes (+0.07%)
2871 bytes +32 bytes (+1.13%) +30 bytes
2904 bytes +65 bytes (+2.29%) +63 bytes
2917 bytes +78 bytes (+2.75%) +76 bytes
2946 bytes +107 bytes (+3.77%) +105 bytes
2981 bytes +142 bytes (+5.00%) +140 bytes
3017 bytes +178 bytes (+6.27%) +176 bytes
3051 bytes +212 bytes (+7.47%) +210 bytes

Non-DEFLATE Algorithms

Archivers based on completely different compression algorithms often produce superior results.
Unfortunately, browsers only support gzip compression at the moment.
Algorithm Program Parameters Size Compared To Best Zopfli
RAR (proprietary) RAR rar a -m5 -md64m -mc63:128t -mt1 2561 bytes -278 bytes (-9.79%)
Brotli (Wikipedia) brotli brotli -q 11 2636 bytes -203 bytes (-7.15%)
ZSTD (Wikipedia) zstd zstd -19 2859 bytes +20 bytes (+0.70%)
PPMd (Wikipedia) 7zip 7za a -mx=9 -m0=ppmd 2876 bytes +37 bytes (+1.30%)
Burrows-Wheeler transform (Wikipedia) bzip2 bzip2 -9 2893 bytes +54 bytes (+1.90%)
LZMA2 (Wikipedia) xz xz -9 2932 bytes +93 bytes (+3.28%)
ZPAQ (Wikipedia) zpaq zpaq -method 69 3437 bytes +598 bytes (+21.06%)

Detailled Analysis

I wrote a DEFLATE decoder in Javascript. Click the button below to start a client-side analysis of the smallest gzipped files (may take a second):


Notes: pigz is a fast open source multi-threaded implementation of gzip written by one of the original authors of gzip.
However, when using compression level 11, pigz actually switches to the slower Zopfli algorithm and isn't multi-threaded anymore.
KrzyMOD's extensions to Zopfli offer the highest level of configuration and is therefore used for my brute-force search.
Ken Silverman wrote the closed-source KZIP compression program and Jonathon Fowler ported it to Linux.
Defluff was created by Joachim Henke; DeflOpt is a tool by Ben Jos Walbeehm.

website made by Stephan Brumme in 2015 and still improving in 2018.
all timestamps are displayed in central european time. see my changelog.
no flash, not even images or external css files - and everything squeezed into a single html file.
which was handsomely compressed before releasing it into the wild internet - obviously.

please visit my homepage and my blog, too.
email: minime (at) stephan-brumme.com