Back to models

phi-3-mini 3.8B

ECI#201
117
Microsoft
Apr. 23, 2024
Open weights

phi-3-mini 3.8B is an open-weights model from Microsoft, released April 23, 2024. It has an ECI score of 117 and ranks 201st of 249 tracked models. At the time of its release, it ranked 19th of 59 tracked models.

Input price—
Output price—
Parameters3.8e9
Training compute7.5e22 FLOP

phi-3-mini 3.8B vs select alternatives

Reasoning
DomainBenchmark
phi-3-mini 3.8B
Microsoft
Apr. 23, 2024
Open weights
Phi-2
Microsoft
Dec. 12, 2023
Open weights
Llama 3-8B
Meta AI
Apr. 18, 2024
Open weights
DeepSeek-V2 (MoE-236B, May 2024)
DeepSeek
May 7, 2024
Open weights
ECI scoreBest score 166
117108116125
Ranking#201#219#205#176
Conceptual reasoning
BBHBest score 89.2%
71.7%
59.4%
n/a
78.8%
Games
Chess PuzzlesBest score 72.0%
0.0%
n/a
0.0%
n/a
Language understanding
Adversarial NLIBest score 58.1%
52.8%
42.5%
57.3%
n/a
WinoGrandeBest score 89.2%
70.8%
54.7%
75.7%
86.3%
HellaSwagBest score 95.3%
76.7%
53.6%
n/a
87.1%
Science
OpenBookQABest score 88.0%
88.0%
73.6%
82.6%
n/a
ARC AI2Best score 95.3%
84.9%
75.9%
82.8%
92.2%
World knowledge
TriviaQABest score 87.5%
64.0%
45.2%
67.7%
80.0%
MMLUBest score 88.1%
68.8%
58.4%
68.8%
78.4%

Model metadata

General
OrganizationMicrosoft
Release dateApr. 23, 2024
Model accessibilityOpen weights
CountryUnited States of America
Knowledge cutoffUnknown
Provider notesPhi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone
Pricing
Input price—
Output price—
Training
Parameters3.8e9
Training compute7.5e22 FLOP
Training hardwareNVIDIA H100 SXM5 80GB
Training time7 days
Training dataset size3.3e12 datapoints