I wonder if having a big mixture of experts isn't all that valuable for the type of tasks in math and coding benchmarks. Like my intuition is that you need all the extra experts because models store fuzzy knowledge in their feed-forward layers, and having a lot of feed-forward weights lets you store a longer tail of knowledge. Math and coding benchmarks do sometimes require highly specialized knowledge, but if we believe the story that the experts specialize to their own domains, it might be that you only really need a few of them if all you're doing is math and coding. So you can get away with a non-mixture model that's basically just your math-and-coding experts glued together (which comes out to about 32B parameters in R1's case).
MoE is likely temporary, local optimum now that resembles bitter lesson path. With the time we'll likely distill what's important, shrink it and keep it always active. There may be some dynamic retrieval of knowledge (but not intelligence) in the future but it probably won't be anything close to MoE.
It would be interesting if research teams would try to collapse trained MoE into JoaT (Jack of all Trades - why not?).
With MoE architecture it should be efficient to back propagate other expert layers to align with result of selected one – at end changing multiple experts into multiple Jacks.
Having N multiple Jacks at the end is interesting in itself as you may try to do something with commonalities that are present, available on completely different networks that are producing same results.
I had a similar confusion previously, so maybe I can help. I used to think that a mixture of experts model meant that you had like 8 separate parallel models, and you would decide at inference time which one to route to. This is not the case, the mixture happens at a much smaller scale.
Instead, the mixture of experts exists within individual layers. Suppose we want to have a big feed-forward layer that takes as input a 1024-element vector, has a hidden size of 8096, and an output size of 1024. We carve up that 8096 hidden layer into 8 1024-sized chunks (this does not have to be the same size as the input). Whenever an input arrives at this layer, a routing function determines which of those 1024-sized chunks should serve as the hidden layer. Every token within a single prompt/response can choose a different chunk when it is processed by this layer, and every layer can have a different routing decision. So if I have 100 layers, each of which has 8 experts, there are 8^100 possible different paths that an individual token could take through the network.
I think it will be more akin to o1-mini/o3-mini instead of r1. It is a very focused reasoning model good at math and code, but probably would not be better than r1 at things like general world knowledge or others.