New: Agent Index for knowledge work is live.Explore the index ↗
Compare models

Best LLMs for precise instruction following

IFBench

IFBench tests whether models follow precise output requirements and generalize to 58 diverse, verifiable constraint types. It measures instruction-following reliability rather than completion of a full agent workflow.

Updated: Data as of:
Widely usedNot saturated
Official benchmark
IFBench results. Each row identifies its benchmark version and evaluation track. Scores from different versions are not directly comparable. Best effort from the latest stored report date per model and version.
Model / creatorVersionReported score
MiniMax M3MiniMax · Open source / open weights58 constraint types82.857%
DeepSeek V4 Flash 0420DeepSeek · Open source / open weights58 constraint types79.184%
Qwen3.5-397B-A17BQwen · Open source / open weights58 constraint types78.8%
Gemini 3 FlashGoogle · Proprietary58 constraint types78%
Gemini 3.1 ProGoogle · Proprietary58 constraint types77.1%
DeepSeek V4 Pro 0424DeepSeek · Open source / open weights58 constraint types76.463%
GLM-5.1Z.ai · Open source / open weights58 constraint types76.3%
Kimi K2.6Moonshot AI · Open source / open weights58 constraint types75.986%
GPT-5.5OpenAI · Proprietary58 constraint types75.9%
Muse SparkMeta · Proprietary58 constraint types75.9%
MiniMax M2.7MiniMax · Open source / open weights58 constraint types75.7%
GLM-5.2Z.ai · Open source / open weights58 constraint types73.333%
Kimi K2.5Moonshot AI · Open source / open weights58 constraint types70.204%
gpt-oss-120bOpenAI · Open source / open weights58 constraint types68.98%
Kimi K2 ThinkingMoonshot AI · Open source / open weights58 constraint types68.095%
gpt-oss-20bOpenAI · Open source / open weights58 constraint types65.102%
DeepSeek R1 (Jan 2025)DeepSeek · Open source / open weights58 constraint types38.98%
View all 17 published reports
MiniMax M382.857%
58 constraint types

selection: best reported configuration per release · MiniMax-M3 · tools: benchmark-defined

58 constraint types

selection: best reported configuration per release · DeepSeek V4 Flash 0420 (Max) · tools: benchmark-defined

58 constraint types

selection: publisher default displayed configuration · Qwen3.5 397B A17B

58 constraint types

selection: publisher default displayed configuration · Gemini 3 Flash

58 constraint types

selection: publisher default displayed configuration · Gemini 3.1 Pro Preview

58 constraint types

selection: best reported configuration per release · DeepSeek V4 Pro 0424 (Max) · tools: benchmark-defined

GLM-5.176.3%
58 constraint types

selection: publisher default displayed configuration · GLM-5.1

Kimi K2.675.986%
58 constraint types

selection: best reported configuration per release · Kimi K2.6 (Reasoning) · tools: benchmark-defined

GPT-5.575.9%
58 constraint types

selection: publisher default displayed configuration · GPT-5.5 (xhigh)

58 constraint types

selection: publisher default displayed configuration · Muse Spark

58 constraint types

selection: publisher default displayed configuration · MiniMax-M2.7

GLM-5.273.333%
58 constraint types

selection: best reported configuration per release · GLM-5.2 (Max) · tools: benchmark-defined

Kimi K2.570.204%
58 constraint types

selection: best reported configuration per release · Kimi K2.5 (Reasoning) · tools: benchmark-defined

58 constraint types

selection: best reported configuration per release · gpt-oss-120b (High) · tools: benchmark-defined

58 constraint types

selection: best reported configuration per release · Kimi K2 Thinking · tools: benchmark-defined

58 constraint types

selection: best reported configuration per release · gpt-oss-20b (High) · tools: benchmark-defined

58 constraint types

selection: best reported configuration per release · DeepSeek R1 (Jan '25) · tools: benchmark-defined