Home › Language Models › Phi-4-mini-instruct

Phi-4-mini-instruct - AI Language Models Tool

Overview

Phi-4-mini-instruct is a lightweight, instruction-tuned language model from Microsoft in the Phi-4 family. At 3.8 billion parameters it is engineered for instruction-following and reasoning workloads where memory, latency, or compute are constrained. The model supports a very long context (128K tokens) and was post-trained with both supervised fine-tuning (SFT) and direct preference optimization (DPO) to improve instruction adherence and alignment for chat-style and tool-enabled function-calling prompts. ([huggingface.co](https://huggingface.co/microsoft/Phi-4-mini-instruct?utm_source=openai)) Microsoft and the Hugging Face model card position Phi-4-mini-instruct for commercial and research use-cases where strong stepwise reasoning, multilingual support, and long-context handling matter — examples include long-document summarization, multi-step reasoning assistants, lightweight on-prem or edge agents, and latency-sensitive API deployments. The model is released under an MIT license and is available via Hugging Face and Microsoft Foundry/Azure delivery channels. Users should still validate outputs for safety and factuality in high-risk settings. ([huggingface.co](https://huggingface.co/microsoft/Phi-4-mini-instruct?utm_source=openai))

Model Statistics

  • Downloads: 224,880
  • Likes: 686
  • Pipeline: text-generation

License: mit

Model Details

Architecture and core specs: Phi-4-mini-instruct is a dense decoder-only Transformer with ~3.8B parameters and a context length of 131,072 (128K) tokens. The model uses architectural efficiency improvements compared with prior Phi-mini models (grouped-query attention and shared input/output embeddings) and exposes a large tokenizer/vocabulary (reported up to ~200,064 tokens). Training details published on the Hugging Face model card list roughly 5 trillion tokens of training data, training run on hundreds of GPUs (documented as 512 A100-80G for the instruct variant) and an offline knowledge cutoff of June 2024; training windows and dates reported span late 2024 with a public release in early 2025. ([huggingface.co](https://huggingface.co/microsoft/Phi-4-mini-instruct?utm_source=openai)) Fine-tuning and alignment: The model received supervised fine-tuning and direct preference optimization (DPO) post-training to improve instruction-following, reduce common failure modes, and increase helpfulness in chat and function-calling settings. The model card also documents safety-red-teaming observations (residual jailbreak risks, function-name/URL hallucination risks in tool-calling scenarios) and recommends domain-specific evaluation before production deployment. ([huggingface.co](https://huggingface.co/microsoft/Phi-4-mini-instruct?utm_source=openai)) Inference and software: Phi-4-mini-instruct is distributed with PyTorch/Transformers support and is compatible with long-context inference toolchains (vLLM, FlashAttention-enabled runtimes). The card notes hardware/testing on GPUs that support FlashAttention and recommends appropriate runtime stacks for efficient 128K token inference. The model is available in multiple community-contributed formats (including quantized builds) to reduce memory/latency for edge or local deployments. ([huggingface.co](https://huggingface.co/microsoft/Phi-4-mini-instruct?utm_source=openai))

Key Features

  • 3.8B-parameter dense decoder-only Transformer tuned for instruction following.
  • Extremely long context support: ~128K tokens for long-document tasks.
  • Post-training alignment via supervised fine-tuning (SFT) and DPO for better instruction adherence.
  • Large tokenizer / vocabulary (reported up to ~200,064 tokens) for multilingual coverage.
  • Optimized for low-latency and memory-constrained deployments; compatible with vLLM and FlashAttention runtimes.
  • Released under an MIT license and available via Hugging Face and Microsoft Foundry/Azure.

Example Usage

Example (python):

### Minimal example: load and generate with the Hugging Face Transformers pipeline
# Note: For production long-context inference (128K) use vLLM or FlashAttention-optimized runtimes.
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import torch

MODEL = "microsoft/Phi-4-mini-instruct"

tokenizer = AutoTokenizer.from_pretrained(MODEL, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    torch_dtype=torch.float16,
    device_map="auto",
)

gen = pipeline("text-generation", model=model, tokenizer=tokenizer, device=model.device)

prompt = "<|system|>You are a concise assistant.<|end|><|user|>Explain the core idea of Fourier transforms in simple terms.<|end|><|assistant|>"
output = gen(prompt, max_new_tokens=180, do_sample=False)
print(output[0]['generated_text'])

# For heavy long-context use prefer vLLM or the model's FlashAttention recommendations in the model card. (See Hugging Face model page.)

Benchmarks

MMLU (aggregate reported by Microsoft Foundry): 67.3% (Source: https://ai.azure.com/catalog/models/Phi-4-mini-instruct)

BigBench Hard (CoT) — reported internal comparison: 70.4% (Source: https://ai.azure.com/catalog/models/Phi-4-mini-instruct)

HumanEval (third-party measurement): 74.3% (Source: https://benched.ai/models/phi-4-mini)

MATH-500 / MATH500 (third-party measurement): 69.6% (Source: https://benched.ai/models/phi-4-mini)

Last Refreshed: 2026-02-24

Key Information

  • Category: Language Models
  • Type: AI Language Models Tool
Visit Official Website

Related Tools

Qwen2.5-7B

Qwen2.5-7B is a large language model designed for text generation, featuring improvements in coding, mathematics, instruction following, long text generation, and multilingual support. It supports context lengths up to 128K tokens and is intended for sophisticated NLP tasks.
DeepSeek-V3

Open‑weight flagship LLM (base and chat) with strong reasoning and tool‑use; runs with SGLang/LMDeploy/vLLM and permissive model license.
Llama 3

Llama 3 is an open access large language model (LLM) released by Meta, available in various configurations (8B and 70B parameters) with capabilities for fine-tuning and integrations into platforms like Hugging Face, Google Cloud, and Amazon SageMaker.
UNfilteredAI-1B

A large-scale text generation model designed for creative and unconstrained content production without traditional filtering.
Shuttle-3

Shuttle-3 is a state-of-the-art language model designed for high-quality text generation, particularly suited for complex chat, multilingual communication, and reasoning tasks. It is fine-tuned from the Qwen-2.5-72b-Instruct model and designed to emulate high-quality prose similar to Claude 3 models.
WizardLM

WizardLM is a state-of-the-art large language model designed for complex chat, multilingual tasks, reasoning, and agent functionalities. It features an AI-powered pipeline (Auto Evol-Instruct) that optimizes instruction datasets for improved performance across various domains and leverages Arena Learning for an expanded learning pool of challenging instruction data.