MMLU Redux

MMLU Redux - Benchmark Leaderboard & Model Performance | AI Stats

Models Using This Benchmark

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

Recorded Results

Average Score

Score Range

Leading Model

Models Using This Benchmark