MMODELYST
Evals/IFBench
EVL

IFBench

benchmarkInstruction Followingaccuracy %
Leaderboard · 100 models
accuracy % — higher is better
1Grok 4.20 0309
82.9
2MiniMax-M3
82.9
3Nemotron 3 Ultra 550B A55B
81.4
4Grok 4.3
81.3
5Grok 4.20 0309 v2
81.2
6Qwen3.7 Max
80.5
7Nemotron Cascade 2 30B A3B
80.4
8MiMo-V2.5-Pro
79.9
9DeepSeek V4 Flash
79.2
10Nova 2.0 Pro Preview
79
11Qwen3.5 397B A17B
78.8
12Gemini 3 Flash Preview
78
13Qwen3.7 Plus
78
14GPT-5.2 Codex
77.6
15Gemini 3.1 Flash-Lite
77.2
16Gemini 3.1 Pro Preview
77.1
17Qwen3.6 Max Preview
76.6
18DeepSeek V4 Pro
76.5
19GLM-5.1
76.3
20Gemini 3.5 Flash
76.3
21Kimi K2.6
76
22Muse Spark
75.9
23GPT-5.4 nano
75.9
24GPT-5.5
75.9
25Qwen3.5 122B A10B
75.7
26MiniMax-M2.7
75.7
27Qwen3.5 27B
75.6
28Gemma 4 31B
75.6
29GPT-5.2
75.4
30GPT-5.3 Codex
75.4
31Qwen3.6 Plus
75.2
32GPT-5 Codex
74.1
33GPT-5.4
73.9
34Command A+
73.9
35Gemma 4 12B
73.5
36GLM-5.2 (max)
73.3
37GPT-5.4 mini
73.3
38GLM-5-Turbo
73.2
39GPT-5
73.1
40GPT-5.1
72.9
41GPT-5.6 Sol (max)
72.7
42Qwen3.5 35B A3B
72.5
43Gemma 4 26B A4B
72.4
44GLM-5
72.3
45MiniMax-M2
72.3
46MiMo-V2-Flash (Feb 2026)
71.8
47MiniMax-M2.5
71.6
48GPT-5.5 Instant (May 2026)
71.5
49NVIDIA Nemotron 3 Super 120B A12B
71.5
50o3
71.4
51GPT-5 mini
71.2
52Solar Pro 3
71.2
53GPT-5.6 Terra (max)
71.2
54NVIDIA Nemotron 3 Nano 30B A3B
71.1
55GPT-5.6 Sol
71
56Qwen3 Max Thinking
70.7
57Gemini 3 Pro Preview
70.4
58o1
70.3
59Kimi K2.5
70.2
60GPT-5.1 Codex
70
61MiniMax-M2.1
69.9
62Mercury 2
69.8
63Apriel-v1.6-15B-Thinker
69.1
64gpt-oss-120b
69
65MiMo-V2-Pro
68.8
66Mistral Medium 3.5
68.8
67o4-mini
68.7
68Nova 2.0 Lite
68.5
69KAT-Coder-Pro V1
68.4
70Kimi K2 Thinking
68.1
71GLM-4.7
67.9
72GPT-5.1 Codex mini
67.9
73GPT-5 nano
67.6
74Qwen3.6 27B
67.6
75MiMo-V2-Omni-0327
67.3
76Step 3.7 Flash
67.3
77MiMo-V2.5
67.1
78Qwen3.5 9B
66.7
79KAT Coder Pro V2
66.7
80HyperNova 60B 2605
66.5
81Step 3.5 Flash 2603
66.5
82GPT-5.6 Terra
66.3
83Nex-N2-Pro
66.2
84Nova 2.0 Omni
66.2
85Olmo 3.1 32B Think
66
86gpt-oss-20b
65.1
87K-EXAONE
64.7
88Step 3.5 Flash
64.6
89Qwen3.6 35B A3B
64.4
90MiMo-V2-Flash
64.2
91DeepSeek V3.2 Speciale
63.9
92Claude Fable 5
63.5
93Nemotron 3 Nano Omni 30B A3B Reasoning
63.2
94Hy3-preview
63.1
95Kimi K2.7 Code
63.1
96K2 Think V2
62.8
97Claude Opus 4.8
62.2
98Apriel-v1.5-15B-Thinker
61.7
99GLM 5V Turbo
61.1
100GLM-4.7-Flash
60.8
Benchmark health
active

Discriminating but climbing — the frontier is moving up this benchmark.

80.9
top-10 mean
82.9
best
100
models
Recent changes
GPT-5.6 Terra (max)+71.26d ago
GPT-5.6 Terra+66.36d ago
GPT-5.6 Sol (max)+72.76d ago
GPT-5.6 Sol+716d ago
GPT-5.5 Instant (May 2026)+71.513d ago
What improves this score

Datasets & environments shown to raise it.

No data yet.
Research using this benchmark
all →
Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B AgentFAPO: Fully Autonomous Prompt Optimization of Multi-Step LLM PipelinesAMARIS: A Memory-Augmented Rubric Improvement System for Rubric-Based Reinforcement Learning

Matched by name in title/abstract.