Azure Updates: Phi-4-mini-flash-reasoning; AI Foundry Agent Service; Azure Accelerate

Microsoft CVP Weizhu Chen shared recent developments with Phi-4-mini-flash reasoning. Microsoft has previously launched multiple installments of the Phi model family. With its new Phi-4-mini-flash-reasoning offering, the company aims to improve latency, compute, and memory for edge devices, mobile apps, and other systems that face resource constraints.

Although the new model shares a name with Phi-4-mini, it is built on a different hybrid architecture with 10 times higher throughput with an average of two to three times average reduction in latency that allows significantly faster inferencing without reducing reasoning performance. Currently, the model is available through AI Foundry, the NVIDIA API Catalog, and Hugging Face. 

“At the core of Phi-4-mini-flash-reasoning is the newly introduced decoder-hybrid-decoder architecture, SambaY, whose central innovation is the Gated Memory Unit (GMU), a simple yet effective mechanism for sharing representations between layers.  The architecture includes a self-decoder that combines Mamba (a State Space Model) and Sliding Window Attention (SWA), along with a single layer of full attention,” Chen wrote.

VP of Product Yin Arenas discussed the introduction of Deep Research in AI Foundry Agent Service. The new offering is an API and software development kit offering based on OpenAI’s advanced agentic research capability that integrates with Azure’s existing agentic platform. With the public preview, developers are able to build agents to analyze and synthesize information from across the web for complex research tasks. By providing Deep Research as an agentic tool Microsoft aims to enable developers to automate web-scale research, programmatically build agents, orchestrate complex workflows, and achieve enterprise governance.