Phi-4-mini-flash-reasoning

Phi-4-mini-flash-reasoning is a 3.8-billion-parameter reasoning model announced by Microsoft in May 2025 and released in June 2025 under the MIT License. It keeps the math-focused reasoning training of Phi-4-mini-reasoning but replaces the transformer with the SambaY decoder-hybrid-decoder architecture.

SambaY combines state-space layers, differential attention, grouped-query attention, a gated memory unit, shared key-value memory, and one global-attention layer, with a 64k-token context window. Microsoft reported up to 10x higher decoding throughput than Phi-4-mini-reasoning in a vLLM test with a 2k prompt and 32k generated continuation, targeting low-latency environments.