Jeffrey Zhou et al.

Instruction-Following Eval

A benchmark of 541 prompts built from 25 verifiable instruction types. It tests whether a model follows checkable constraints such as keyword, length, casing, and response-format requirements.

In index
Instruction
Source ↗

Models33

Top result95.0%Qwen3.5-27B

Top-3 spread0.2 pts

Year2023

Result and price

Results

33 results

#ModelResult

Qwen3.5-27B95.0%

02INAgents-A194.8%

03INAgents-A1-4B94.8%

Qwen3.7 Plus94.6%

Qwen3.7 Max94.3%

Qwen3.6 Plus94.3%

07DSdots3-note Preview93.9%

07MAKimi K2.593.9%

o3-mini93.9%

Qwen3.5-122B-A10B93.4%

GLM-592.6%

Qwen3.5 397B92.6%

13LAK-EXAONE 2.092.4%

o192.2%

Qwen3.5-35B-A3B91.9%

16LILFM2.5-8B-A1B91.8%

17PMTernary Bonsai 2 27B91.3%

Claude Opus 4.590.9%

GPT-4.1 mini88.5%

GPT-4.187.4%

21OPMiniCPM5-2B86.7%

DeepSeek V386.1%

23ZYZAYA1-8B85.6%

GPT-4.1 nano83.2%

25LILFM2.5-VL-3B82.3%

26KAKanana-2 3B Instruct81.0%

27CECeleris-180.8%

28OPMiniCPM5-1B80.4%

29KAKanana-2 1.3B Instruct77.6%

30JEMellum2-12B-A2.5B-Thinking76.5%

31JEMellum2-12B-A2.5B-Instruct75.8%

32LILFM2.5-230M71.7%

33LILFM2.5-VL-450M61.2%