Phi-3.5 MoE Instruct

Phi-3.5 MoE Instruct is Microsoft's first mixture-of-experts model, released August 2024 with 41.9B total parameters across 16 experts but only ~6.6B active parameters per forward pass. That active-param footprint gives it a favorable cost profile on inference hardware while hitting reasoning benchmark scores closer to dense 14B models than to pure 6B baselines. The 131K context window carries over from the Phi-3 line. Hosted coverage is thinner than Llama or Qwen equivalents — Azure AI is the primary route, with a handful of generalist aggregators rounding out the options. MIT license removes any commercial friction. If MoE active-param economics matter for your deployment cost model, this is worth benchmarking against the Qwen MoE tier.