Back to models

Phi-2

ECI#219
108
Microsoft
Dec. 12, 2023
Open weights

Phi-2 is an open-weights model from Microsoft, released December 12, 2023. It has an ECI score of 108 and ranks 219th of 249 tracked models. At the time of its release, it ranked 17th of 39 tracked models.

Input price—
Output price—
Parameters2.7e9
Training compute2.3e22 FLOP

Phi-2 vs select alternatives

Reasoning
DomainBenchmark
Phi-2
Microsoft
Dec. 12, 2023
Open weights
Phi-1.5
Microsoft
Sep. 11, 2023
Open weights
Yi 6B
01.AI
Nov. 22, 2023
Open weights
Mixtral 8x7B
Mistral AI
Dec. 11, 2023
Open weights
ECI scoreBest score 166
10891104118
Ranking#219#243#226#197
Conceptual reasoning
BBHBest score 89.2%
59.4%
n/a
47.2%
n/a
Language understanding
Adversarial NLIBest score 58.1%
42.5%
n/a
n/a
55.2%
WinoGrandeBest score 89.2%
54.7%
73.4%
71.3%
77.2%
HellaSwagBest score 95.3%
53.6%
47.6%
74.4%
86.7%
Science
OpenBookQABest score 88.0%
73.6%
37.2%
n/a
85.8%
ARC AI2Best score 95.3%
75.9%
44.4%
50.3%
87.3%
World knowledge
TriviaQABest score 87.5%
45.2%
n/a
n/a
82.2%
MMLUBest score 88.1%
58.4%
37.6%
64.0%
70.6%

Model metadata

General
OrganizationMicrosoft
Release dateDec. 12, 2023
Model accessibilityOpen weights
CountryUnited States of America
Knowledge cutoffUnknown
Provider notesPhi-2: The surprising power of small language models
Pricing
Input price—
Output price—
Training
Parameters2.7e9
Training compute2.3e22 FLOP
Training hardwareNVIDIA A100
Training time14 days
Training dataset size2.5e11 datapoints