# Phi-4 ![Organization](https://img.shields.io/badge/Org-Microsoft-0078d4) ![Parameters](https://img.shields.io/badge/Params-14B-1f6feb) ![Context](https://img.shields.io/badge/Context-16K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Apache%202.0-2da44e) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) ![Released](https://img.shields.io/badge/Released-Dec%202024-555) > **Quick answer:** Phi-4 is Microsoft's December 2024 14B-parameter language model, scoring 70.4% on MMLU-Pro, 75.4% on Arena Hard, 82.6% on HumanEval, and 27.4% on BigCodeBench. Under Apache 2.0 license with open weights, it delivers GPT-4-class knowledge at a 14B parameter scale. The model uses a 16K-token context window. ## At a Glance **Where Phi-4 leads** - 70.4% MMLU-Pro — competitive with models 3-5× its parameter count - 75.4% Arena Hard — strong general instruction following - 82.6% HumanEval — solid function-level code completion - Apache 2.0 — unrestricted commercial use - 14B parameters — highly efficient for self-hosting - Trained on synthetic high-quality data — disproportionately capable for its size **Where it lags** - 16K context window — significantly smaller than Qwen3 32B (131K) or Llama 4 models - 27.4% BigCodeBench — moderate performance on complex code generation - Text-only: no vision capability - 14B dense — no MoE efficiency benefits **Best for:** Edge deployment, resource-constrained inference, self-hosting on consumer GPUs, and tasks requiring GPT-4-class knowledge at 14B scale. ## What Phi-4 Is Phi-4 is Microsoft's December 2024 small language model, the fourth generation of the Phi series. Microsoft's Phi research focuses on demonstrating that high-quality training data — rather than raw parameter count — can yield outsized capability in small models. Phi-4 at 14B achieves MMLU-Pro scores competitive with much larger open-weight models. The model is trained primarily on synthetic data: carefully curated textbook-quality content, code, and reasoning problems. This approach produces a model that excels at the academic knowledge and reasoning tasks covered by MMLU-Pro and Arena Hard, with 70.4% and 75.4% scores respectively placing it well above other models in its size class. Phi-4's 16K context window is its most significant limitation relative to peers: Qwen3 32B offers 131K context at similar size, and Llama 4 Scout offers 10M. For tasks requiring long context, alternative open-weight models are recommended. For tasks fitting within 16K tokens, Phi-4 offers exceptional capability at 14B. ## Specifications | Field | Value | |---|---| | Organization | Microsoft | | Parameters | 14B (dense) | | Context window | 16,384 tokens | | License | Apache 2.0 | | HuggingFace | [microsoft/phi-4](https://huggingface.co/microsoft/phi-4) | | Release date | December 12, 2024 | | Knowledge cutoff | June 2024 | | Modality | Text only | | Architecture | Dense transformer | ## Pricing Phi-4 is available as open weights under Apache 2.0 — free to self-host. Available via Azure AI Foundry and third-party providers. ## Context Window Phi-4 has a **16,384-token context window** — approximately **12 pages** of text. This is significantly smaller than most 2025-era models; for tasks requiring longer context, consider Qwen3 32B (131K) or Llama 4 Maverick (1M). ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [MMLU-Pro](/benchmarks/mmlu-pro) | 70.4% | Benchgen evaluation | 2025-07 | | [Arena Hard](/benchmarks/arena-hard-v2) | 75.4% | Benchgen evaluation | 2025-07 | | [HumanEval](/benchmarks/humaneval) | 82.6% | Benchgen evaluation | 2025-07 | | [BigCodeBench](/benchmarks/bigcodebench) | 27.4% | Benchgen evaluation | 2025-07 | ## Phi-4 vs Alternatives | Model | MMLU-Pro | HumanEval | Context | License | |---|---|---|---|---| | **Phi-4** | 70.4% | 82.6% | 16K | Apache 2.0 | | [Qwen3 32B](/models/qwen/qwen3-32b) | — | — | 131K | Apache 2.0 | | [Llama 4 Scout](/models/meta/llama-4-scout) | 74.3% | — | 10M | Llama 4 | | [Hermes 3 70B](/models/nousresearch/hermes-3-70b) | 47.2% | — | — | Llama 3.1 | Phi-4's 70.4% MMLU-Pro at 14B is a standout result — Llama 4 Scout at 74.3% requires 17B active MoE parameters with a much larger infrastructure footprint. For tasks within 16K context, Phi-4 is the highest-knowledge-per-parameter open-weight model in its class. ## Run Phi-4 ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "microsoft/phi-4" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto") inputs = tokenizer("Explain Newton's third law:", return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=512) print(tokenizer.decode(outputs[0], skip_special_tokens=True)) ``` ## Related Models & Benchmarks - Compare: [Qwen3 32B](/models/qwen/qwen3-32b) · [Llama 4 Scout](/models/meta/llama-4-scout) · [Hermes 3 70B](/models/nousresearch/hermes-3-70b) - Benchmarks: [MMLU-Pro](/benchmarks/mmlu-pro) · [Arena Hard](/benchmarks/arena-hard-v2) · [HumanEval](/benchmarks/humaneval) ## Frequently Asked Questions

What is Phi-4?

Phi-4 is Microsoft's December 2024 14B-parameter language model, scoring 70.4% MMLU-Pro and 75.4% Arena Hard under Apache 2.0. It achieves GPT-4-class knowledge at 14B scale through synthetic high-quality data training.

Is Phi-4 open source?

Yes. Phi-4 is released under Apache 2.0 on Hugging Face — free for commercial use with no restrictions.

What is Phi-4's context window?

Phi-4 has a 16,384-token context window. For longer context tasks, consider Qwen3 32B (131K) or Llama 4 Maverick (1M).

What makes Phi-4 different from other small models?

Phi-4 is trained primarily on synthetic high-quality data (textbook-quality content, curated code) rather than raw web data. This produces a model that punches above its weight on knowledge and reasoning benchmarks — 70.4% MMLU-Pro at 14B parameters is comparable to much larger models.

--- *Specs from Microsoft's official Phi-4 release (December 2024) and Benchgen evaluations. Last updated 2026-07-24.*

Phi 4 — Evaluations

Evaluations