MMLU Redux

AI benchmark results and model performance

Type: percentage

Recorded Results

Average Score

81.47%

Score Range

43.20% - 94.90%

Leading Model (lowest score)

43.20% - Ernie 4.5 21B A3B

Scores Over Time

Individual benchmark scores plotted by date.

Models Using This Benchmark

Organisation	Model	Reported	Top Score	Info	Self Reported	Source
Baidu	Ernie 4.5 21B A3B	-	43.20%	inferred version-family alias from ernie-4.5	Yes	Source
Baidu	Ernie 4.5 21B A3B Thinking	-	43.20%	inferred version-family alias from ernie-4.5	Yes	Source
Baidu	Ernie 4.5 300B A47B	-	43.20%	inferred version-family alias from ernie-4.5	Yes	Source
Baidu	Ernie 4.5 Turbo	-	43.20%	inferred version-family alias from ernie-4.5	Yes	Source
Baidu	Ernie 4.5 VL 28B A3B	-	43.20%	inferred version-family alias from ernie-4.5	Yes	Source
Baidu	Ernie 4.5 VL 424B A47B	-	43.20%	inferred version-family alias from ernie-4.5	Yes	Source
Qwen	Qwen 3.5 0.8B	02 Mar 2026	59.50%	-	Yes	Source
Qwen	Qwen 2.5 Coder 3B	-	71%	inferred family alias from qwen2.5-omni-7b (score=0.3000; benches=45)	Yes	Source
Qwen	Qwen 2.5 Coder 7B	-	71%	inferred high-confidence family alias from qwen2.5-omni-7b (score=0.4700; benches=45)	Yes	Source
Qwen	Qwen 2.5 Math 7B	-	71%	inferred high-confidence family alias from qwen2.5-omni-7b (score=0.4767; benches=45)	Yes	Source
Qwen	Qwen 2.5 Math 7B PRM800K	-	71%	inferred family alias from qwen2.5-omni-7b (score=0.3696; benches=45)	Yes	Source
Qwen	Qwen 2.5 Math PRM 7B	-	71%	inferred family alias from qwen2.5-omni-7b (score=0.4092; benches=45)	Yes	Source
Qwen	Qwen 2.5 Omni 3B	-	71%	inferred high-confidence family alias from qwen2.5-omni-7b (score=0.4933; benches=45)	Yes	Source
Qwen	Qwen 2.5 Omni 7B	-	71%	-	Yes	Source
LG	EXAONE 4.0 1.2B	15 Jul 2025	71.50%	Reasoning	Yes	Source
Mistral	Ministral 3.0 3B	02 Dec 2025	73.50%	5 Shot	Yes	Source
Qwen	Qwen 7B	-	75.40%	inferred family alias from qwen-2.5-7b-instruct (score=0.3083; benches=14)	Yes	Source
Mistral	Ministral 3.0 8B	02 Dec 2025	79.30%	5 Shot	Yes	Source
Qwen	Qwen 3.5 2B	02 Mar 2026	79.60%	-	Yes	Source
Qwen	Qwen 14B	-	80%	inferred family alias from qwen-2.5-14b-instruct (score=0.3060; benches=16)	Yes	Source
Qwen	Qwen 3 VL 4B Instruct	-	81.50%	-	Yes	-
Mistral	Ministral 3.0 14B	02 Dec 2025	82%	5 Shot	Yes	Source
Qwen	Qwen 3 VL 8B Instruct	-	84.90%	-	Yes	-
Qwen	Qwen 3 4B Instruct 2507	-	86%	inferred family alias from qwen3-vl-4b-thinking (score=0.3462; benches=48)	Yes	-
Qwen	Qwen 3 4B SafeRL	-	86%	inferred family alias from qwen3-vl-4b-thinking (score=0.3850; benches=48)	Yes	-
Qwen	Qwen 3 4B Thinking 2507	-	86%	inferred family alias from qwen3-vl-4b-thinking (score=0.3462; benches=48)	Yes	-
Qwen	Qwen 3 Embedding 4B	-	86%	inferred family alias from qwen3-vl-4b-thinking (score=0.3850; benches=48)	Yes	-
Qwen	Qwen 3 Guard Gen 4B	-	86%	inferred family alias from qwen3-vl-4b-thinking (score=0.3400; benches=48)	Yes	-
Qwen	Qwen 3 Guard Stream 4B	-	86%	inferred family alias from qwen3-vl-4b-thinking (score=0.3371; benches=48)	Yes	-
Qwen	Qwen 3 Reranker 4B	-	86%	inferred family alias from qwen3-vl-4b-thinking (score=0.3850; benches=48)	Yes	-
Qwen	Qwen 3 VL 4B Thinking	-	86%	-	Yes	-
Qwen	Qwen 72B	-	86.80%	inferred family alias from qwen-2.5-72b-instruct (score=0.3060; benches=14)	Yes	Source
Qwen	Qwen 3 235B A22B	-	87.40%	-	Yes	Source
Qwen	Qwen 3 VL 30B A3B Instruct	-	88.40%	-	Yes	-
Qwen	Qwen 3 8B	-	88.80%	inferred high-confidence family alias from qwen3-vl-8b-thinking (score=0.4600; benches=50)	Yes	-
Qwen	Qwen 3 Guard Stream 8B	-	88.80%	inferred family alias from qwen3-vl-8b-thinking (score=0.3371; benches=50)	Yes	-
Qwen	Qwen 3 VL 8B Thinking	-	88.80%	-	Yes	-
Qwen	Qwen 3 VL Embedding 8B	-	88.80%	inferred high-confidence family alias from qwen3-vl-8b-thinking (score=0.5232; benches=50)	Yes	-
Qwen	Qwen 3 VL Reranker 8B	-	88.80%	inferred high-confidence family alias from qwen3-vl-8b-thinking (score=0.5275; benches=50)	Yes	-
Qwen	Qwen 3.5 4B	02 Mar 2026	88.80%	-	Yes	Source
DeepSeek	DeepSeek OCR	20 Oct 2025	89.10%	inferred family alias from deepseek-v3 (score=0.3000; benches=20)	Yes	Source
DeepSeek	DeepSeek V2 (2024-06-28)	28 Jun 2024	89.10%	inferred family alias from deepseek-v3 (score=0.4159; benches=20)	Yes	Source
DeepSeek	DeepSeek V4	-	89.10%	inferred high-confidence family alias from deepseek-v3 (score=0.5818; benches=20)	Yes	Source
Qwen	Qwen 3 VL 32B Instruct	-	89.80%	-	Yes	-
Qwen	Qwen 3 Next 80B A3B Instruct	-	90.90%	-	Yes	Source
Qwen	Qwen 3 VL 30B A3B Thinking	-	90.90%	-	Yes	-
Qwen	Qwen 3.5 9B	02 Mar 2026	91.10%	-	Yes	Source
DeepSeek	DeepSeek V3.1	21 Aug 2025	91.80%	Non-thinking: 91.8%, Thinking: 93.7%	Yes	Source
DeepSeek	DeepSeek V3.1 Terminus	22 Sept 2025	91.80%	inferred alias from deepseek-v3.1	Yes	Source
Qwen	Qwen 3 VL 32B Thinking	-	91.90%	-	Yes	-
Qwen	Qwen 3 VL 235B A22B Instruct	-	92.20%	-	Yes	-
LG	EXAONE 4.0 32B	15 Jul 2025	92.30%	Reasoning	Yes	Source
Qwen	Qwen 3 Next 80B A3B Thinking	-	92.50%	-	Yes	Source
Moonshot	Kimi K2 (2025-09-05)	05 Sept 2025	92.70%	EM	Yes	Source
Qwen	Qwen 3 A235 A22B Instruct 2507	-	93.10%	-	Yes	Source
Qwen	Qwen 3.5 27B	24 Feb 2026	93.20%	-	Yes	Source
Qwen	Qwen 3.5 Flash	23 Feb 2026	93.20%	inferred family alias from qwen3.5-27b (score=0.4147; benches=81)	Yes	Source
Qwen	Qwen 3.5 35B A3B	24 Feb 2026	93.30%	-	Yes	Source
Qwen	Qwen 3 VL 235B A22B Thinking	-	93.70%	-	Yes	-
Qwen	Qwen 3 235B A22B Thinking 2507	-	93.80%	-	Yes	Source
Qwen	Qwen 3.5 122B A10B	24 Feb 2026	94%	-	Yes	Source
Moonshot	Kimi K2 Thinking	06 Nov 2025	94.40%	inferred alias from kimi-k2-thinking-0905	Yes	Source
Qwen	Qwen 3.6 Plus	01 Apr 2026	94.50%	-	Yes	Source
Qwen	Qwen 3.5 397B A17B	16 Feb 2026	94.90%	-	Yes	Source