Ziim, 12201222 bytes

In an unexpected turn of events, I'm counting this submission in UTF-16.

↘ ↓↘↘↘↘↙↓↘↘ ↙↓↘ ↓↘ ↙↘←↓ ↘↘↙↓↘↘↘↙↓↘↘↘↘↙↓↘↘↘↘↙↓ ↘↘↙↓↘↘↙↓↘ ↙ ↘ ↓↘ ↙↘←↓↘ ↙ ↘ ↓ ↘↘↙↓↘↘↙↓↘↘↘↘↙↓ ↘ ↙↘←↓↘ ↙ ↘ ↓ ↘ ↙ ↘ ↓↘↘↙↓↘↘↘↙↓↘↘↘↙↓↘↘ ↙↓↘←↓↘ ↙↓↘ ↓↘↘↙↓ ↘↘↙↓↘↘↙↓↘ ↙↘←↓↘ ↙↘←↓ ↘ ↙↘←↓ ↘↘↙↓↘↘↙↓ ↘ ↙ ↘ ↓↘↘ ↙ ↘ ↓↘ ↙↓↘←
↓→↘ ↑↑↓→↘ ↑↓  ↓→↘ ↓↓↓→↘↘ ↓→↘ ↑↓→↘ ↑↑↓→↘ ↑↑↓→↘↘ ↓→↘ ↓→↘ ↓↓↓↓→↘ ↓↓↓→↘ ↓↓↓↓→↘↘ ↓→↘ ↓→↘ ↑↑↓→↘↘ ↓↓↓→↘ ↓↓↓↓→↘↘ ↓↓↓↓→↘ ↓→↘ ↑↓→↘ ↑↓→↘ ↑↓  ↓→↘ ↓  ↓→↘ ↓→↘↘ ↓→↘ ↓→↘ ↓↓↓→↘ ↓↓↓→↘↘ ↓↓↓→↘↘ ↓→↘ ↓→↘↘ ↓↓↓↓→↘ ↑↓↓↓↓→↘ ↓  ↓→
↗  ↑↖ ↗  ↑↖↗↑↗↗  ↑↗↗↗  ↑↑↗  ↑↖↗  ↑↖ ↗  ↑↖ ↗  ↑↑↗  ↑↗  ↑↗↗↗↗  ↑↗↗↗  ↑↗↗↗↗  ↑↑↗  ↑↗  ↑↖ ↗  ↑↑↗↗↗  ↑↗↗↗↗  ↑↑↗↗↗↗  ↑↗  ↑↖↗  ↑ ↗  ↑↖↗↑↗↗  ↑↗↑↗↗  ↑↗  ↑↑↗  ↑↗  ↑↗↗↗  ↑↗↗↗  ↑↑↗↗↗  ↑↑↗  ↑↗  ↑↑↗↗↗↗  ↑↖↗↗↗↗  ↑↗↑↗↗

Because the code snippet above will likely not render in a very readable manner for you (either because of missing characters, or because the diagonal arrows aren't as monospaced as they should be), here is an image of the solution. I've marked the concatenators where a new byte starts in red:

^{Click for larger version}

Well, this is a bit embarrassing. I spent quite some time on this solution, and at 610 characters, it is 267 characters shorter than the one on the esolangs page. However, it contains a lot more arrows, and because those are 3 bytes in UTF-8, it's actually longer than the one on esolangs (1516 vs 1345). But because this solution is so dense and contains far more arrows than spaces, I'm actually saving bytes by using UTF-16 and counting each character as two bytes.

Anyway, regardless of how we're counting, this is far from optimal. The above code is entirely linear, and builds one bit at a time. So I'm not reusing any bitstrings (like I could for the ls or os or even other substrings). I chose this approach because I wanted to start with something where it would be easier to pack the arrows very densely. If one used a more square layout, one could make use of a lot more repetition in the output (on a bit level), and thereby make massive savings.