Ai2 EMO MoE clusters by domain

This is the kind of MoE release that actually changes the routing conversation: if the specialization holds up, document-level gating could make MoEs easier to interpret and more useful on real workloads, not just benchmarks.

• –The 1B-active setup keeps inference relatively cheap while preserving the capacity of a much larger 14B model.
• –Domain-shaped experts suggest the router is learning higher-level structure, which is more promising than pure surface-form clustering.
• –If this generalizes, it could improve long-form coherence and reduce expert thrashing on mixed-topic documents.
• –The tradeoff is obvious: stronger inductive bias can help specialization, but it may hurt flexibility on short, heterogeneous prompts.
• –Open availability on Hugging Face makes EMO a good comparison point against token-routed MoEs and Ai2's earlier OLMoE-style work.