microsoft/phi-4

Weights: 29.3 GB · safetensors
Source: huggingface page
Commit: master

Evaluated across 12 benchmarks. Ranks in the top 3 on 0 of 12. Strongest showing on EQ-Bench (77.6 eqbench, #14 of 28). Weakest on BBH (0.4% exact_match, #23 of 28).

Benchmark results

Benchmark	Metric	Score	Rank	Actions
EQ-Bench	eqbench	77.6±1.6	#14 of 28	View run →
MMLU	acc	76.3%±0.3%	#10 of 28	View run →
MBPP	pass@1	72.2%±2.0%	#4 of 28	View run →
MMLU-Pro	exact_match	71.8%±0.4%	#7 of 28	View run →
IFEval	prompt_level_strict_acc	62.7%±2.1%	#23 of 28	View run →
GPQA Diamond	acc	35.9%±3.4%	#6 of 28	View run →
GPQA Extended	acc	34.2%±2.0%	#6 of 28	View run →
GPQA Main	acc	33.5%±2.2%	#8 of 28	View run →
LongBench	aggregate	31.7%±0.4%	#6 of 17	View run →
MGSM	exact_match	23.8%±0.7%	#27 of 28	View run →
GSM8K	exact_match	9.7%±0.8%	#24 of 28	View run →
BBH	exact_match	0.4%±0.1%	#23 of 28	View run →

EQ-Bench#14 of 28
eqbench
77.6±1.6
View run →
MMLU#10 of 28
acc
76.3%±0.3%
View run →
MBPP#4 of 28
pass@1
72.2%±2.0%
View run →
MMLU-Pro#7 of 28
exact_match
71.8%±0.4%
View run →
IFEval#23 of 28
prompt_level_strict_acc
62.7%±2.1%
View run →
GPQA Diamond#6 of 28
acc
35.9%±3.4%
View run →
GPQA Extended#6 of 28
acc
34.2%±2.0%
View run →
GPQA Main#8 of 28
acc
33.5%±2.2%
View run →
LongBench#6 of 17
aggregate
31.7%±0.4%
View run →
MGSM#27 of 28
exact_match
23.8%±0.7%
View run →
GSM8K#24 of 28
exact_match
9.7%±0.8%
View run →
BBH#23 of 28
exact_match
0.4%±0.1%
View run →

last evaluated:7 months ago

How to cite

Citation

FrozeBench. "microsoft/phi-4." https://frozebench.com/models/microsoft-phi-4. Retrieved 2026-06-04.

BibTeX

@misc{frozebench_microsoft_phi_4,
  title = {microsoft/phi-4},
  howpublished = {\url{https://frozebench.com/models/microsoft-phi-4}},
  year = {2026},
  note = {FrozeBench. Retrieved 2026-06-04.}
}

URL

https://frozebench.com/models/microsoft-phi-4