Does anyone use the Zillions Of Games program?

http://www.zillions-of-games.com

I'm tempted to sue them for false advertising! Zillions? Pah, more like a 50 or so. But still, it has many chess variations available. Not just xiangqi and shogi but variations on Western chess. The playing strength isn't all that strong, but then having a program that plays so many games well is definitely too good to be true. I

It also counts the games that you can implement using it. Not only the games that others happen to have implemented with it for you. So I don't think there is anything misleading here.

That being said, I don't use Zillions. The scope of Zillions is very wide (including cards and dice), and I am mostly interested in proper Chess variants. So I prefer to use WinBoard, in combination with configurable Chess-variant engines like Sjaak II or Fairy-Max. The latter alone already supports 30 western Chess variants. Sjaak II supports an even wider range of Chess variants, including those with drops, such as Shogi, and zoned boards, such as Xiangqi. These engines are in general also stonger than Zillions. And for games with very special rules (such as Chu Shogi) there often are specialized engines available.

Just curious, how strong is it in e.g. ELO terms?

That hasn't really been measured, because it doesn't support a standard protocol for chess engines (like CECP or UCI), and thus cannot play automated games. And for a reasonably accurate rating determination you need many hundreds of games.

In addition, it would not be equally strong in all variants it can play. It has a heuristic for estimating piece values, and this heuristic is known to frequently fail in predicting reasonable values. It does not seem to understand divergent pieces, (with separate moves for capture and non-capture), so for pieces that can be dropped (like in Shogi) is uses ridiculously large values.

Its search is also not very stable: in variants with too powerful pieces it suffers from search explosion. There are several implementations of Tenjiku Shogi in it, but none of those can finish the 1-ply iteration in the initial position even in 10 min, and as a consequence it then plays the first move of the static move ordering. (Which sacrifices a Bishop General on the Side Soldier.

For orthodox Chess and 10x8 Capablanca variants it probably plays reasonably well, both search-wise and evaluation-wise, but even then it tends to lose from Fairy-Max (which for normal Chess is the same as micro-Max, rated 1868 Elo at CCRL), in the few operator-mediated games these played against each other.

I wish alpha zero was open source it technically could play any variant

There does exist an open-source implementation of the Alpha-Zero algorithm, called Leela Zero (for Go) or Leela Chess Zero ('LC0') for orthodox Chess.

Of course you would have to adapt the source code if you want it to play a variant (board size and move generator, possibly other rules such as winning condition). And then train it. Which for Leela/Alpha Zero is no small effort; Google had its globe-spanning network of servers to do it, for LC0 it has been a massive group effort of people donating CPU time.

Did anyone realise how weak the engine in chessvariants.com for atomic is? It can not even see a mate in 4!

I suppose you are referring to the Interactive Diagram that goes with the article about Atomic Chess. Then it really should not come as a surprise. The I.D. there is designed as a sparring partner for readers who have just acquainted themselves with the rules, still giving them a good chance for beating it. You don't want to turn away people from a variant by clobbering them on their first experience.

So by default it is set to search 2.5 ply ahead, adjustable between 1 and 4 ply through buttons below the Diagram. And actually ply is already pretty challenging. But to see mate in 4 requires a look ahead of 9 ply. Most humans would not be able to calculate that deep, and would have no chance at all against a program that searches 9 ply.