ApX logoApX logo

Phi-4 Reasoning Plus

Parameters

14B

Context Length

33K

Modality

Text

Architecture

Dense

License

MIT

Release Date

30 Apr 2025

Knowledge Cutoff

Mar 2025

System Requirements

VRAM requirements for different quantization methods and context sizes

1,024 tokens

31.12 GB VRAM

Consumer

2x RTX 4090

24GB VRAM

Datacenter

1x NVIDIA A100

80GB VRAM

Apple Silicon

1x Apple M3 Max

128GB VRAM

32,768 tokens

37.95 GB VRAM

Consumer

2x RTX 4090

24GB VRAM

Datacenter

1x NVIDIA A100

80GB VRAM

Apple Silicon

1x Apple M3 Max

128GB VRAM

Architecture Diagram

Input TokensToken EmbeddingPosition: AbsoluteHidden: 5.1k · Context: 33K · Vocab: 100.4kx 40 layersRMSNormPre-AttentionMulti-Head Attention40Q / 10KV headsHead dim: 128+RMSNormPre-FFNFeed-Forward NetworkSwiGLUIntermediate: 17.9k+Final RMSNormOutput Logits

Evaluation Benchmarks

Rank

#164

BenchmarkScoreRank

Professional Knowledge

MMLU Pro

0.76

60

Rankings

Overall Rank

#164

Coding Rank

-

About Phi-4 Reasoning Plus

Phi-4 Reasoning Plus is a 14-billion parameter language model engineered by Microsoft to provide advanced chain-of-thought processing and high-precision logical inference. As an enhanced variant in the Phi-4 family, it is designed to handle sophisticated problem-solving across domains such as mathematics, scientific inquiry, and complex code generation. The model produces structured outputs that include an explicit reasoning trace followed by a final solution, facilitating transparency in its decision-making process. This design prioritizes output quality and depth for tasks where thoroughness is more critical than immediate response speed.

Technically, the model utilizes a dense, decoder-only Transformer architecture with multi-head attention (MHA). It incorporates Rotary Position Embeddings (RoPE) and an expanded context window of 32,768 tokens, allowing it to maintain coherence over the lengthy sequences often required for multi-step reasoning. The training methodology represents a significant advancement in data-centric AI, employing supervised fine-tuning (SFT) on over 1.4 million chain-of-thought traces, followed by reinforcement learning using the Group Relative Policy Optimization (GRPO) algorithm. This RL phase specifically targets verifiable mathematical and logical problems, refining the model's ability to self-correct and explore alternative solutions.

Operational characteristics of Phi-4 Reasoning Plus include a notable increase in token generation compared to the standard Phi-4 models, as the 'plus' variant typically produces 50% more tokens to provide more exhaustive explanations. While this results in higher latency, it enables the model to rival the performance of much larger systems in specialized benchmarks. The model is released under the MIT license with open weights, making it accessible for deployment on consumer-grade hardware and local environments where computational resources are constrained but high-fidelity reasoning is required.

Technical Specifications

Attention

Attention Structure

Multi-Head Attention

Attention Heads

40

Key-Value Heads

10

Attention Head Dimension

-

Position Embedding

Absolute Position Embedding

RoPE Theta

500,000

Sliding Window Attention

No

Sliding Window Size

-

Sliding Window Ratio

-

Linear Attention

-

Linear Attention Ratio

-

Normalization

RMS Normalization

Activation Function

SwigLU

Dimensions

Hidden Dimension Size

5,120

Number of Layers

40

FFN Intermediate Size (Dense)

17,920

Multi-Token Prediction Heads

-

Tokenizer

Vocabulary Size

100,352

About Phi-4

The Microsoft Phi-4 model family comprises small language models prioritizing efficient, high-capability reasoning. Its development emphasizes robust data quality and sophisticated synthetic data integration. This approach enables enhanced performance and on-device deployment capabilities.


Other Phi-4 Models
Phi-4 Reasoning Plus: Specifications and GPU VRAM Requirements