Phi-4 Reasoning Plus

Microsoft provides high transparency regarding the model's lineage, explicitly identifying it as a fine-tuned variant of the Phi-4 base model. The architecture is documented as a dense, decoder-only Transformer with 14 billion parameters, utilizing Multi-Head Attention (MHA) and Rotary Position Embeddings (RoPE). The technical report details the transition from the base Phi-4 to the 'Reasoning' and 'Reasoning Plus' variants, including the specific addition of <think> and </think> tokens and the expansion of the context window to 32,768 tokens. The use of Group Relative Policy Optimization (GRPO) for the reinforcement learning phase is also clearly stated.

The training methodology is described as 'data-centric,' with specific disclosures about the SFT phase using 1.4 million chain-of-thought traces generated via o3-mini and the RL phase using ~6,000 high-quality math problems. While the report mentions a blend of synthetic data and filtered public domain data (web, code, STEM), it lacks a precise percentage breakdown of the pre-training data composition (e.g., exact ratios of web vs. books vs. code). However, the disclosure of the specific 'teacher' models (o3-mini, DeepSeek-R1 for the mini variant) and the data-cleansing methodology (decontamination in Appendix B) provides better-than-average transparency.

The model uses the tiktoken tokenizer with a stated vocabulary size of 100,352 tokens. Documentation confirms the inclusion of specific reserved tokens for reasoning traces (<think>, </think>). The tokenizer is publicly accessible via the Hugging Face repository and is integrated into the standard 'transformers' library (version 4.51.3+), allowing for direct verification of tokenization behavior and alignment with the claimed language support (primarily English).

The parameter count is consistently stated as 14 billion. As a dense architecture, the active parameters equal the total parameters, which is explicitly confirmed in technical documentation to avoid MoE-related confusion. Detailed architectural specs, including the context window (32K) and the specific modifications from the Phi-3/Phi-4 base, are well-documented in the technical report.

Microsoft provides specific hardware and compute metrics for the reasoning variants: the training utilized 32 H100-80G GPUs over a duration of 2.5 days. While the carbon footprint is not explicitly calculated in the model card, the disclosure of GPU hours and hardware type allows for independent estimation. This level of detail is significantly higher than the industry standard for 'open-weight' models.

The technical report includes comprehensive evaluations on standard benchmarks (AIME, GPQA, MATH, LiveCodeBench) and compares results against both open and proprietary models. Microsoft specifies the use of the 'simple-evals' framework for reproducibility and provides details on the prompting strategy (temperature=0.8, top_p=0.95). However, while the methodology is described, the full evaluation code and exact prompt sets for every benchmark are not always provided in a single, turn-key repository.

The model is designed with a specific 'reasoning' identity, producing structured outputs with clear reasoning traces. It identifies as a Microsoft Phi-family model and maintains version consistency across its different variants (Mini, Reasoning, Reasoning Plus). There are no documented instances of the model claiming to be a competitor's system (e.g., GPT-4), and its limitations regarding English-only support and math-specialization are clearly disclosed.

The model is released under the MIT license, which is a highly permissive, standard open-source license. The terms are clear, allowing for both commercial and non-commercial use, and there are no conflicting proprietary 'community licenses' often seen with other large providers. The weights are freely available for download on Hugging Face.

VRAM requirements are well-documented, with specific guidance provided for different hardware (e.g., 28GB for full precision, fits on 2x RTX 4090 or 1x A100). Microsoft also provides optimized ONNX versions with 4-bit quantization (RTN) and documents the resulting performance/latency trade-offs. The impact of the 'Plus' variant's longer reasoning traces on latency is explicitly mentioned.

Microsoft uses clear naming conventions (Phi-4-reasoning vs. Phi-4-reasoning-plus) and maintains a changelog on the Hugging Face repository. However, as a relatively new release (April 2025), there is limited long-term data on how Microsoft handles silent updates or model drift over time. The model is currently described as 'static,' which aids transparency but requires monitoring for future 'silent' revisions.