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

Phi-4-mini-instruct

4-miniย ยทย Family: Phi
Microsoft's open Phi-4 family language model (~3.8B parameters) with a 128K context window and function calling support; MIT license.
โœ“ Activeโœ“ Public accessโš– Open sourceLLM๐Ÿ“ Phi
Context window
128K
tokens
Parameters
3.8B
parameters
Release date
26 February 2025
๐ŸขMicrosoftProducer
Access:DownloadAPIHostedDeployment:๐Ÿ’ป Localโ˜ Cloud๐Ÿ“ฑ On-device

Overview

Phi-4-mini-instruct is an open, dense decoder-only language model from Microsoft's Phi-4 family with about 3.8B parameters. It uses a Transformer architecture with grouped-query attention (GQA), shared input/output embeddings and SwiGLU activation.

It supports a 128K token context window and a tokenizer with a 200,064-token vocabulary. The model was trained on about 5 trillion tokens (public data cutoff June 2024). It supports function calling and multilingual processing. Released under the MIT license.

Classification
LLM
Family: Phi
Access & deployment
DownloadAPIHosted
LocalCloudOn-device
Weights: Open source
Key parameters
๐Ÿ“ Context: 128K
๐Ÿงฉ Parameters: 3.8B
โœ“ Toolsย ยทย โœ“ Fine-tuning
๐Ÿ“ฅ Input: text

Technical specification

Context window
128K
tokens
Parameters
3.8B
parameters
Knowledge cutoff
1 Jun 2024
Knowledge boundary
License
MIT
Hardware requirements
About 3.8B parameters; runs on a single GPU, and the ONNX variant enables inference on edge/CPU devices.
Features:โœ“ Tool useโœ“ Fine-tuning
Modalities
โฌ‡ Input
text
โฌ† Output
textcode

Capabilities and applications

Native model capabilities
Language modeling
Ability to predict subsequent tokens and generate coherent natural-language text based on the preceding context.
Category: language
Reasoning
The model's ability to reason logically and solve complex problems.
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
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language
Long context
Support for large context windows โ€” tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Instruction following
Precisely following instructions contained in the prompt: response format, length, style, constraints (e.g. 'reply in six words'). GPT-5.1 significantly improved this capability compared to GPT-5.
Category: language

Benchmark results

10 benchmarks
MMLU
accuracy ยท 5-shot
67.3%
๐Ÿ“„ Hugging Face (karta modelu)
MMLU-Pro
accuracy ยท 0-shot, CoT
52.8%
๐Ÿ“„ Hugging Face (karta modelu)
GSM8K
accuracy ยท 8-shot, CoT
88.6%
๐Ÿ“„ Hugging Face (karta modelu)
MATH
accuracy ยท 0-shot, CoT
64.0%
๐Ÿ“„ Hugging Face (karta modelu)
MGSM
accuracy ยท 0-shot, CoT
63.9%
๐Ÿ“„ Hugging Face (karta modelu)
GPQA
accuracy ยท 0-shot, CoT
25.2%
๐Ÿ“„ Hugging Face (karta modelu)
ARC-Challenge
accuracy ยท 10-shot
83.7%
๐Ÿ“„ Hugging Face (karta modelu)
HellaSwag
accuracy ยท 5-shot
69.1%
๐Ÿ“„ Hugging Face (karta modelu)
BigBench Hard
accuracy ยท 0-shot, CoT
70.4%
๐Ÿ“„ Hugging Face (karta modelu)
Arena Hard
score ยท win rate
32.8%
๐Ÿ“„ Hugging Face (karta modelu)

Technical architecture

Sources and related pages