Phi 3.5 MoE

Phi-3.5-MoE is a small, efficient MoE LLM from Microsoft. It routes tokens across 16 experts (~3.8B each) with top-2 gating (~6.6B active parameters), delivering solid reasoning and code performance at low latency. The model is trained on high-quality, reasoning-dense data and post-trained with SFT plus RL-style techniques (e.g., PPO/DPO) for instruction following and safety. It supports a 128K context window, multilingual use, and an MIT open-weight license. Builders can deploy it serverlessly in Azure AI or run/fine-tune locally (e.g., LoRA). Typical uses include long-context QA and summarization, coding assistants, math, and multilingual applications in constrained environments.