Microsoft Research unveiled Phi‑4 Reasoning on April 30, 2025, alongside two sister models—Phi‑4‑Mini‑Reasoning (≈3.8 B parameters) and Phi‑4‑Reasoning‑Plus (14 B parameters with reinforcement learning tuning). Unlike general‑purpose LLMs, these models are specialized for reasoning: they allocate additional inference compute to verify and refine each solution step. Training leveraged high‑quality web data, synthetic problem sets, and curated “chain‑of‑thought” demonstrations from […]