LANGUAGE MODEL Microsoft Last updated:

Phi-4

Microsoft's Synthetic-Data Small Model

Microsoft's December 2024 small model — only 14 billion parameters but matching GPT-4o on reasoning benchmarks. The trick: Phi-4 was trained mostly on SYNTHETIC data — examples generated by a larger model and filtered for quality. It's a proof point that data quality matters more than raw web-scraped quantity.

Intelligence
Below avg
Speed
Slow
38 tok/s output
Cost
Low
$0.13 in / $0.50 out
Context
16K
Up to 16,384 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on Artificial Analysis. Cost buckets are fixed dollar thresholds keyed off output-token price ($/M out).
Intelligence
  • Top 1%≤ 1%
  • Top 5%≤ 5%
  • Top 10%≤ 10%
  • Good≤ 25%
  • Medium≤ 50%
  • Below avg> 50%
Speed
  • Top 1%≥ 345 tok/s
  • Top 5%≥ 237 tok/s
  • Top 10%≥ 196 tok/s
  • Good≥ 146 tok/s
  • Medium≥ 90 tok/s
  • Slow< 90 tok/s
Cost
  • Freeopen weights · self-host
  • Low< $1 / M out
  • Moderate$1–5 / M out
  • High≥ $5 / M out

Why it matters

Validated synthetic-data training at scale. Showed that closed labs' web-scraping advantage matters less than once thought — a small model + great data can beat a huge model + average data.

Core Capabilities

Long Documents
Handles entire codebases, books, and multi-doc RAG.
Generative
Produces images, video, audio, or other media.
Research
Foundational paper or scientific contribution.

Context Window

16k tokens
≈ short doc
4k Chat 聊天
32k Long docs 长文档
128k Books 整本书
400k Multi-doc 多文档
1M Codebase 整个代码库
10M
16k

Availability

API
Not available
Product / App
Not available
Open Source
Released
Enterprise

Pricing Model

Free / self-host
Open weights — pay only for compute.
Self-host

Capability / Performance

Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).

Lower 20% Upper 80% This model
Reasoning
AA Intelligence Index · scaled to 10
1.7
5.6
1.5
Coding
SciCode · scaled to 10
1.8
4.3
2.6
Agentic tasks
Terminal-Bench Hard · scaled to 10
0.2
3.6
0.4
Context / memory
Context window size · log-scaled
6.0
9.0
3.0
Cost efficiency
Input price ($/M tokens) · cheaper scores higher
6.2
10.0
10.0
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.

What it feels like

  • 14B parameters that outperforms models 5x larger on math and reasoning — synthetic-data scaling laws in action
  • MMLU 84.8% (up from Phi-3's 77.9%); MATH 56.1% (up from 42.5%) — large quality jumps without size jump
  • HumanEval 82.6% — competitive with much larger frontier-tier models on Python code generation
  • Phi-4-reasoning variants beat o1-mini and DeepSeek-R1-Distill-Llama-70B on most reasoning benchmarks
  • Phi-4-reasoning-plus surpasses full DeepSeek-R1 on AIME 2025 — 14B beating 671B is the headline
  • Trained on 9.8T tokens over three weeks, mixing GPT-4o-generated synthetic data with curated web text

Best use cases

  • On-device reasoning where larger models can't fit
  • Math, science, and code benchmarks at SLM cost
  • Self-hosted production where 14B inference is affordable on a single GPU
  • Distillation experiments and SLM research baselines

Tools to try

Not ideal for

  • Broad-knowledge factual QA — small models still have limited recall
  • Multimodal tasks (Phi-4 base is text-only; vision lives in Phi-4-multimodal)
  • Workloads needing 100K+ context (limited window vs frontier models)

Model Evolution

phi is Microsoft's language model family.

View full evolution tree →

Microsoft Research