Robots Atlas>ROBOTS ATLAS
Phi-4
AI Modelsโ€บPhi

Phi-4

4ย ยทย Family: Phi
Microsoft's open (MIT) small language model: 14B parameters (dense, decoder-only), text-only, 16K context; trained partly on synthetic data, strong at reasoning, math and code.
โœ“ Activeโœ“ Public accessโš– Open sourceLLMReasoning model๐Ÿ“ Phi
Context window
16K
tokens
Parameters
14B (dense, decoder-only)
parameters
Release date
12 December 2024
๐ŸขMicrosoftProducer
Access:DownloadAPIHostedDeployment:๐Ÿ’ป Localโ˜ Cloud

Overview

Phi-4 is a small language model (SLM) developed by Microsoft Research, released on December 12, 2024. It is a dense, decoder-only Transformer with 14B parameters, text-only (best suited for English-language prompts), with a 16K-token context.

Phi-4 continues the Phi family's data-centric strategy: it was built on a blend of synthetic data, filtered public-domain data and acquired datasets, with an emphasis on high quality and advanced reasoning. As a result, despite its small size, it achieves strong results on STEM tasks - especially math (MATH), science questions (GPQA) and coding.

The model underwent fine-tuning and alignment (including supervised fine-tuning and preference optimization). Weights are released under the MIT license (fully open source, self-hosting), and the model is available via Hugging Face, Azure AI Foundry and provider APIs. Trained on 1920 H100 GPUs over 21 days (9.8T tokens).

Classification
LLMReasoning model
Family: Phi
Applications
CodingQ&A / Question answeringResearch assistanceKnowledge workChatbots
Access & deployment
DownloadAPIHosted
LocalCloud
Weights: Open source
Key parameters
๐Ÿ“ Context: 16K
๐Ÿงฉ Parameters: 14B (dense, decoder-only)
โœ“ Fine-tuning
๐Ÿ“ฅ Input: text
Platforms
Hugging Face HubOpenRouter

Technical specification

Context window
16K
tokens
Parameters
14B (dense, decoder-only)
parameters
License
MIT
Hardware requirements
14B - self-hosting on a single GPU (~28 GB VRAM in BF16, or quantization on smaller GPUs). Supported: Transformers, vLLM, llama.cpp, Ollama. Training: 1920x H100-80G, 21 days.
Features:โœ“ Fine-tuning
Modalities
โฌ‡ Input
text
โฌ† Output
textcode

Capabilities and applications

Native model capabilities
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Mathematical reasoning
The model's ability to solve mathematical tasks requiring multi-step reasoning โ€” equations, proofs, combinatorics, geometry, calculus and competition-level problems.
Category: reasoning
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Language modeling
Ability to predict subsequent tokens and generate coherent natural-language text based on the preceding context.
Category: language
Application domains
CodingQ&A / Question answeringResearch assistanceKnowledge workChatbots

Pricing

Technical architecture

Core Architecture
TRTransformer
Model Form
LLLLM
Training Techniques
ITInstruction TuningKDKD

Deployment and security

๐Ÿ”’ Security / Enterprise
โœ“ Verified enterprise information

Open-source model under MIT - full freedom of self-hosting / on-premises with data control. Also available in Azure AI Foundry with Microsoft/Azure security controls.

Updated: 31 Jul 2026โ†— Security documentation

Sources and related pages