Microsoft open-sources Phi-4-reasoning-plus: A 14B model beating 70B rivals

Microsoft’s Phi-4-reasoning-plus is an open-weight 14B dense model purpose-built for deep, structured reasoning in math, science, and coding.

Despite its compact footprint, it punches way above its weight. On the demanding AIME 2025 math exam, its pass@1 accuracy across all 30 questions outperforms the 70B DeepSeek-R1 distillation model and closely trails the massive 671B DeepSeek-R1.

How it was built

Microsoft prioritized data quality over massive volume. The model was trained on just 16 billion tokens, with around 8.3 billion unique tokens drawn from synthetic datasets and curated web sources.

During supervised fine-tuning, Microsoft trained the model to structure its thoughts using special <think> and </think> tokens. This explicitly separates scratchpad reasoning from the final answer, keeping long-form logic coherent.

The team followed up with Group Relative Policy Optimization (GRPO) reinforcement learning using only about 6,400 math problems. The reward setup balances correctness with conciseness and penalizes repetition, coaxing the model into longer, deeper deliberation when it lacks confidence.

Under the hood, it is a dense decoder-only Transformer with a default 32,000 token context window, staying stable up to 64,000 tokens in Microsoft's tests.

How to run it

The model drops under the permissive MIT license, meaning unrestricted commercial use, distillation, and fine-tuning are fully supported.

Ecosystem support is ready on day one. It works directly with Hugging Face Transformers, vLLM, llama.cpp, and Ollama, making it easy to run on local rigs or enterprise servers.

Microsoft recommends using it in a chat setting with a system prompt instructing the model to think step-by-step before answering.

Note that Microsoft frames this model as a specialized reasoning engine and research asset rather than an out-of-the-box conversational bot for generic tasks.

Why it matters

For engineers running inference on local workstations or edge servers, running 70B models is often too slow or resource-heavy. Getting top-tier math and reasoning performance at 14B changes local agent development entirely.

It also highlights how far curated synthetic reasoning traces and lightweight RL have come. With a clean MIT license, Phi-4-reasoning-plus gives developers a reliable workhorse for private distillation and domain-specific fine-tuning.