IFBench leaderboard

IFBench evaluates precise instruction-following generalization on 58 challenging, verifiable out-of-domain constraints. Unlike IFEval which tests familiar constraint types, IFBench specifically measures how well models follow novel instructions they haven't been optimized for, exposing overfitting to common instruction patterns.

Models ranked by their published Instruction Following Benchmark value

Rank	Model	Provider	IFBench
1	MAI-Thinking-1	Microsoft	85
2	Qwen3.8 Max	Alibaba	82.8
3	TMInkling-Small	Thinking Machines Lab	82.2
4	Nemotron 3 Ultra	NVIDIA	81.7
5	Qwen3.8-Omni-Flash	Alibaba	81.5
6	Grok 4.3	xAI	81.3
6	Qwen3.8-Flash-Next	Alibaba	81.3
8	DSdots3-note Preview	Dots Studio	80.4
9	Solar Open 2	Upstage	80
10	TMInkling	Thinking Machines Lab	79.8
11	Qwen3.8-27B	Alibaba	79.5
12	Granite 4.2 8B	IBM	79.33
13	Qwen3.7 Max	Alibaba	79.1
13	Qwen3.7 Plus	Alibaba	79.1
15	Granite 4.2 30B	IBM	77.17
16	Mercury 2.5	Inception	77
16	Muse Glimmer 30B	Meta	77
18	Gemini 3.5 Flash	Google	76.3
19	STA.X K2	SK Telecom	75.9
20	Qwen3.6 Plus	Alibaba	75.8
21	Ling 3.0 Flash	InclusionAI	74.5
22	Granite 4.2 3B	IBM	74.33
23	Nemotron 3 Nano Omni 30B A3B	NVIDIA	74.2
24	PMTernary Bonsai 2 27B	Prism ML	74
25	Ling 3.0 Flash FP8	InclusionAI	73.4
26	Nemotron 3.5 Lightning 30B A3B NVFP4	NVIDIA	72.88
27	K-EXAONE 2.0	LG AI Research	72.6
28	Agents-A1-4B	InternScience	69.1
29	OPMiniCPM5-2B	OpenBMB	66.3
30	Hy3 Preview	Tencent	63.1
31	LFM2.5-2.6B	LiquidAI	59.17
32	Claude Opus 4.5	Anthropic	58
33	Ling 2.6 Flash	InclusionAI	57
34	LFM2.5-8B-A1B	LiquidAI	56.47
35	Solar Pro 3	Upstage	55.78
36	ZYZAYA1-8B	Zyphra	52.56
37	OPMiniCPM5-1B	OpenBMB	46.67
38	LFM2.5-230M	LiquidAI	38.4
39	KAKanana-2 1.3B Instruct	Kakao	34.69
40	KAKanana-2 3B Instruct	Kakao	33.33
41	LFM2.5-VL-3B	LiquidAI	25.8
42	LLaDA2.2-mini	InclusionAI	24.93

Evidence key: Observed

Rows are ordered by the value BenchLM published, highest first. The source does not state whether a higher value is the better result, so this page does not either: for a benchmark that measures a rate of failure, read the table from the bottom.