You can see the Magic Lantern HDMI results here: https://vimeo.com/52822486
Some have said above that ML's implementation is lacking. That seems to misplace the blame on ML. They are working with what they have in terms of the current limitations. The primary limitations are: 1) The resolution that the Canon firmware makes available to HDMI (1620x1080); 2) The fact that the HDMI out is 1080i.
Because of 1, the same frame that is internally recorded must be cropped (zoomed) from the HDMI output, so the resulting resolution is 1620x911 (and does not even appear to perfectly resolve that resolution, just as the native 5D3 footage does not resolve the full 1920x1080 resolution).
The interlacing is a real pain and limitation. While there could be a limitation in the Hyperdeck Shuttle that I used in recording the interlaced signal, I simply could not reliably reconstruct the 24p signal. I could reconstruct it with a lot of effort, but it was a bit soft. Even the 30p signal was a bit soft after deinterlacing.
My own testing using a resolution chart shows that the HDMI out from ML does seem to resolve about the same resolution as the internal recording, albeit with pillar bars (since it is only 1620x1080). Oddly, though, the shot must be framed differently, since you must zoom in. Thus, with a very good deinterlacer, this might be useful for greenscreen work (assuming you don't need the full 16x9 frame).
In short, while the ML clean HDMI out might be useful for some purposes, it doesn't work for me right now in its current implementation to yield any significant benefits. There is a chance that ML can improve their HDMI out. However, I suspect that the 1080i HDMI output is something that they simply cannot overcome.
I don't think the ML implementation gives us any indication of what Canon's firmware will be. If Canon were going to do what ML did, they could have released clean HDMI already. This gives me hope that they will 1) give us at least as good a resolution as currently is recorded internally; and 2) provide a true 24p signal or at least have the 1080i embed the 24p information so that it is easily and cleanly recoverable. If they do these things, then the externally recorded material should show significant benefits, even if it is 8 bits. The benefits will be in less compression artifacts on highly detailed and dynamic (in motion) material from the higher available bit rates and less color compression from 4:2:2 vs. 4:2:0.