Phi-4-reasoning

Phi-4-Reasoning is a 14-billion-parameter model trained on top of the base Phi-4 through supervised fine-tuning, using over 1.4 million carefully curated prompts and high-quality responses generated by the o3-mini1 model. The model specializes in tasks requiring complex, multi-step reasoning and demonstrates the ability to generate detailed reasoning chains while efficiently utilizing computational resources during inference.

The model’s performance metrics are impressive: it outperforms significantly larger open models such as DeepSeek-R1-Distill-Llama-70B and approaches the performance level of the full DeepSeek-R1 model across various benchmarks. On the AIME 2025 math problems, Phi-4-Reasoning shows an improvement of more than 50 percentage points compared to the base Phi-4, and on LiveCodeBench programming tasks, it improves by over 25 percentage points. A notable feature of Phi-4-Reasoning is its ability to generalize and transfer knowledge to tasks not explicitly included in the training data, with improvements also observed on general tasks, including instruction following and toxic content detection.

Phi-4-Reasoning is ideally suited for applications that demand reliable logical reasoning under limited computational resources—such as educational platforms for solving mathematical problems, automated programming systems, scientific research tools, and any applications where a balance between high-quality reasoning and resource efficiency is essential.

Announce Date: 30.04.2025
Parameters: 15B
Context: 33K
Layers: 40
Attention Type: Full Attention
Developer: Microsoft
Transformers Version: 4.51.1
License: MIT

Use our pre-built public endpoints for free to test inference and explore Phi-4-reasoning capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.