MMODELYST
Evals/MMLU-Pro
EVL

MMLU-Pro

Harder multitask QA

Source
benchmarkKnowledge & QAAccuracy (% correct, 10 options)12K items
What this measures

A harder, cleaned-up MMLU: more reasoning-heavy questions with 10 answer choices instead of 4, so lucky guessing matters less. A better separator of strong models.

Sample
A reasoning-heavy version of an MMLU question with 10 options (A–J) where several look plausible and you must actually work it out.
Leaderboard · 100 models
Accuracy (% correct, 10 options) — higher is better
1Gemini 3 Pro Preview
89.8
2Claude Opus 4.5
89.5
3Gemini 3 Flash Preview
89
4Claude 4.1 Opus
88
5MiniMax-M2.1
87.5
6Claude 4.5 Sonnet
87.5
7GPT-5.2
87.4
8Claude 4 Opus
87.3
9GPT-5
87.1
10GPT-5.1
87
11Grok 4
86.6
12GPT-5 Codex
86.5
13DeepSeek V3.2 Speciale
86.3
14DeepSeek V3.2
86.2
15Gemini 2.5 Pro
86.2
16GPT-5.1 Codex
86
17Gemini 2.5 Pro Preview (Mar' 25)
85.8
18GLM-4.7
85.6
19Grok 4.1 Fast
85.4
20Doubao Seed Code
85.4
21o3
85.3
22DeepSeek V3.1 Terminus
85.1
23Grok 4 Fast
85
24DeepSeek V3.2 Exp
85
25DeepSeek R1 0528 (May '25)
84.9
26Cogito v2.1
84.9
27Kimi K2 Thinking
84.8
28DeepSeek R1 (Jan '25)
84.4
29Qwen3 235B A22B 2507
84.3
30MiMo-V2-Flash
84.3
31Claude 4 Sonnet
84.2
32Gemini 2.5 Flash Preview (Sep '25)
84.2
33Qwen3 Max
84.1
34o1
84.1
35Qwen3 Max (Preview)
83.8
36K-EXAONE
83.8
37Claude 3.7 Sonnet
83.7
38Gemini 2.5 Pro Preview (May' 25)
83.7
39Qwen3 VL 235B A22B
83.6
40GLM-4.5
83.5
41DeepSeek V3.1
83.3
42Gemini 2.5 Flash
83.2
43o4-mini
83.2
44ERNIE 5.0 Thinking Preview
83
45Nova 2.0 Pro Preview
83
46GLM-4.6
82.9
47Hermes 4 - Llama-3.1 405B
82.9
48Qwen3 235B A22B 2507 Instruct
82.8
49Qwen3 235B A22B
82.8
50Grok 3 mini Reasoning
82.8
51GPT-5 mini
82.8
52Llama 3.1 Nemotron Ultra 253B v1
82.5
53Qwen3 Max Thinking (Preview)
82.4
54Kimi K2
82.4
55Qwen3 Next 80B A3B
82.4
56Qwen3 VL 235B A22B Instruct
82.3
57Ling-1T
82.2
58INTELLECT-3
82.2
59MiniMax-M2
82
60GPT-5 (ChatGPT)
82
61GPT-5.1 Codex mini
82
62Kimi K2 0905
81.9
63DeepSeek V3 0324
81.9
64Qwen3 Next 80B A3B Instruct
81.9
65Qwen3 VL 32B
81.8
66EXAONE 4.0 32B
81.8
67MiniMax M1 80k
81.6
68Seed-OSS-36B-Instruct
81.5
69GLM-4.5-Air
81.5
70Magistral Medium 1.2
81.5
71Llama Nemotron Super 49B v1.5
81.4
72Mi:dm K 2.5 Pro Preview
81.3
73KAT-Coder-Pro V1
81.3
74Nova 2.0 Lite
81.3
75Hermes 4 - Llama-3.1 70B
81.1
76Mi:dm K 2.5 Pro
80.9
77Nova 2.0 Omni
80.9
78Llama 4 Maverick
80.9
79MiniMax M1 40k
80.8
80Gemini 2.5 Flash-Lite Preview (Sep '25)
80.8
81gpt-oss-120b
80.8
82Qwen3 VL 30B A3B
80.7
83Mistral Large 3
80.7
84Ring-1T
80.6
85GPT-4.1
80.6
86Qwen3 30B A3B 2507
80.5
87Gemini 2.0 Pro Experimental (Feb '25)
80.5
88Solar Pro 2
80.5
89GPT-4o (March 2025, chatgpt-4o-latest)
80.3
90Gemini 2.5 Flash Preview
80
91GLM-4.6V
79.9
92Grok 3
79.9
93Qwen3 32B
79.8
94Gemini 2.0 Flash Thinking Experimental (Jan '25)
79.8
95Motif-2-12.7B-Reasoning
79.6
96DeepSeek R1 Distill Llama 70B
79.5
97NVIDIA Nemotron 3 Nano 30B A3B
79.4
98Grok Code Fast 1
79.3
99Ring-flash-2.0
79.3
100Qwen3 Omni 30B A3B
79.2
Benchmark health
saturating

Top models cluster near the ceiling — differences here are getting too small to mean much. Weight newer, harder benchmarks more.

88.0
top-10 mean
89.8
best
100
models
Recent changes
GPT-5 mini83.7 → 82.811d ago
Nova 2.0 Lite81.8 → 81.318d ago
What improves this score

Datasets & environments shown to raise it.

No data yet.