Skip to main content

Phi-4 Mini Instruct

MicrosoftPhiOpen WeightMIT · Commercial OK

Description

Phi 4 Mini Instruct is a lightweight (3.8B parameters) open model built upon synthetic data and filtered web data, focusing on high-quality reasoning. It supports a 128K token context length and is enhanced for instruction adherence and safety via supervised fine-tuning and direct preference optimization.

Release Date
2024-02-26
Parameters
3.8B
Context Length
128K
Modalities
text

Capability Radar

16
general
12
coding
18
reasoning
20
scienceest.
16
agents
0
multimodal

Science uses a reasoning proxy when dedicated science benchmarks are unavailable.

Rankings

Domain#RankScoreSource
Code Ranking435
11.0
AA
General Ranking481
15.0
AA
Math Reasoning310
18.0
AA
Reasoning54
69.0
LS
Science452
17.0
AA

Benchmark Scores (LLM Stats)

Biology

GPQA25.2%SR

Creativity

Social IQa72.5%SR
Arena Hard32.8%SR

Finance

MMLU67.3%SR
TruthfulQA66.4%SR
MMLU-Pro52.8%SR

General

ARC-C83.7%SR
OpenBookQA79.2%SR
PIQA77.6%SR
Multilingual MMLU49.3%SR

Language

BoolQ81.2%SR
BIG-Bench Hard70.4%SR
Winogrande67.0%SR

Math

GSM8k88.6%SR
MATH64.0%SR
MGSM63.9%SR

Reasoning

HellaSwag69.1%SR

AA Evaluation Indices

Math Index
6.7
Intelligence Index
3.0
Math 500
0.7
Mmlu Pro
0.5
Gpqa
0.3
Ifbench
0.2
Lcr
0.1
Livecodebench
0.1
Scicode
0.1
Tau2
0.1
Aime 25
0.1
Hle
0.0
Aime
0.0
Terminalbench Hard
0.0

LLM Stats Category Scores

Math
70
Psychology
70
Reasoning
70
Language
60
Legal
60
Finance
60
General
60
Healthcare
60
Physics
50
Creativity
50
Biology
30
Chemistry
30
Writing
30

Pricing

Input PriceFree
Output PriceFree
Blended Price (3:1)Free
Cache Read Price$0.08 / 1M tokens

Speed

Tokens/sec46.4
Time to First Token1.21s
Time to Answer1.21s

Provider Price Ranking

Provider Price Ranking

4 providers

Cheapest: Azure Cognitive ServicesMost Expensive: NanoGPT
ProviderInputOutput
1Azure Cognitive ServicesCheapest
$0.075
$0.3
2Azure
$0.075
$0.3
3Weights & Biases
$0.08
$0.35
4NanoGPT
$0.17
$0.68

Compare pricing across different API providers for this model.

External Sources