Phi-4-reasoning-plus

Phi-4-reasoning-plus is a dense 14-billion-parameter language model in Microsoft's Phi-4 family, designed as a reasoning specialist rather than a general chat model. Its architecture follows a classic transformer recipe: 40 layers with a hidden size of 5,120, 40 query attention heads grouped under 10 key-value heads, a head dimension of 128, and a SwiGLU feed-forward block with an intermediate size of 17,920, all normalized through RMSNorm. Absolute positional embeddings are paired with rotary position embeddings using a RoPE theta of 500,000, giving the model a stable sense of token order across long sequences. A 100,352-token vocabulary and an attention layout built for 40 query heads against 10 KV heads point to a design that prioritizes thorough per-token reasoning while keeping key-value memory lean, an arrangement that fits workloads where each step builds on prior context.

This variant is a reinforcement-learning refinement of the base Phi-4 reasoning model, trained on a blend of synthetic and curated public data with an emphasis on math, science, and code. Compared with its sibling, it produces on average about 50% more tokens per response, yielding longer chain-of-thought blocks followed by concise summarization, a pattern that lets the model work through multi-step problems before delivering a final answer. That post-training lineage shows up in third-party rankings where the model is highlighted for strong math performance and respectable showings in legal, finance, and healthcare reasoning, while coding remains a relative weak spot. With MIT-licensed open weights, a 128K-token extended context, and support for tool use, Phi-4-reasoning-plus is well suited for self-hosted assistants that need verifiable, step-by-step reasoning in domains like quantitative analysis, scientific Q&A, and structured problem solving.