Yes, the 3.6 27B is incredible compared to what came before at that size range but, it's worth noting that the benchmarks make it look a lot closer to DS4-Flash than it actually is because fewer param models generalize worse outside what they were post-trained for.

The reason to scale to 3T is not to flex but because scaling total param count is still the best known way to increase model raw intelligence. With MoEs, we can scale effective capacity without paying such a hefty cost on active compute. And recall that due to how effectively attention compounds information, the residual stream is overwhelmed and stops being able to properly add new information at deeper layers.

At larger sizes, MoEs also have some advantages: their residual stream is much less noisy, the conditional computation effectively enables higher per layer pattern compression thanks to specialization (think of a kind of a more domain specific efficient encoding because lots of unrelated features do not need to share the same representational subspace).

Final and most important is not to think of them as just say, 30B of intelligence but as 30B worth of specialized compute dynamically composed per token, combinatorially selected from a (likely redundant but still, combinatorial) library of functions. In fact, over a large sequence, an MoE traces out a trajectory that covers far more than just 30B worth of activated parameters and implicitly composes over the many specialized functions across steps (think of dynamically assembling a program chopped up into 30B sized modules and compute per step).