# Best LLMs for Instruction Following — September 2026 Leaderboard

> As of September 2026, MAI-Thinking-1 leads BenchLM's instruction following leaderboard with a weighted score of 94.7.

- **Last verified:** September 4, 2026
- Canonical page: https://benchlm.ai/instruction-following
- **Ranking coverage:** 124 category-ranked models from 411 tracked models
- **Category weight:** 5% of the overall BenchLM score

## Current ranking

| Rank | Model | Creator | Weighted score | Published category rows | Exact-source rows (all categories) |
|------|-------|---------|----------------|----------------|-------------------|
| 1 | [MAI-Thinking-1](/models/mai-thinking-1) | Microsoft | 94.7 | 1 | 13 total |
| 2 | [Grok 4.3](/models/grok-4-3) | xAI | 94.4 | 2 | 20 total |
| 3 | [GPT-5.2-Codex](/models/gpt-5-2-codex) | OpenAI | 93.5 | 1 | 10 total |
| 4 | [MiMo-V2.5-Pro](/models/mimo-v2-5-pro) | Xiaomi | 93.5 | 1 | 20 total |
| 5 | [MiniMax M3](/models/minimax-m3) | MiniMax | 93.5 | 1 | 33 total |
| 6 | [GLM-5.1](/models/glm-5-1) | Z.AI | 93.5 | 1 | 31 total |
| 7 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 92.9 | 1 | 44 total |
| 8 | [Muse Spark](/models/muse-spark) | Meta | 92.9 | 1 | 31 total |
| 9 | [GPT-5.4 nano](/models/gpt-5-4-nano) | OpenAI | 92.9 | 1 | 25 total |
| 10 | [MiniMax M2.7](/models/minimax-m2-7) | MiniMax | 92.7 | 1 | 30 total |
| 11 | [Qwen3.5-122B-A10B](/models/qwen3-5-122b-a10b) | Alibaba | 92.7 | 2 | 22 total |
| 12 | [Gemma 4 31B](/models/gemma-4-31b) | Google | 92.6 | 1 | 15 total |
| 13 | [Qwen3.5-27B](/models/qwen3-5-27b) | Alibaba | 92.6 | 2 | 21 total |
| 14 | [GPT-5.3 Codex](/models/gpt-5-3-codex) | OpenAI | 92.3 | 1 | 15 total |
| 15 | [GPT-5.2](/models/gpt-5-2) | OpenAI | 92.3 | 1 | 21 total |
| 16 | [Qwen3.7 Max](/models/qwen3-7-max) | Alibaba | 91.1 | 3 | 47 total |
| 17 | [Qwen3.7 Plus](/models/qwen3-7-plus) | Alibaba | 91.1 | 3 | 61 total |
| 18 | [Qwen3.8 Max](/models/qwen3-8-max) | Alibaba | 90.7 | 1 | 55 total |
| 19 | [GPT-5.4](/models/gpt-5-4) | OpenAI | 90.4 | 1 | 41 total |
| 20 | [Command A+](/models/command-a-plus) | Cohere | 90.4 | 1 | 12 total |
| 21 | [Gemma 4 12B](/models/gemma-4-12b) | Google | 89.8 | 1 | 20 total |
| 22 | [GLM-5.2](/models/glm-5-2) | Z.AI | 89.6 | 1 | 30 total |
| 23 | [GPT-5.4 mini](/models/gpt-5-4-mini) | OpenAI | 89.6 | 1 | 26 total |
| 24 | [Inkling-Small](/models/inkling-small) | Thinking Machines Lab | 89.6 | 1 | 32 total |
| 25 | [GLM-5-Turbo](/models/glm-5-turbo) | Z.AI | 89.4 | 1 | 8 total |
| 26 | [GPT-5.1](/models/gpt-5-1) | OpenAI | 89.1 | 1 | 12 total |
| 27 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 88.8 | 1 | 47 total |
| 28 | [Nemotron 3 Ultra](/models/nemotron-3-ultra) | NVIDIA | 88.7 | 2 | 26 total |
| 29 | [Qwen3.5-35B-A3B](/models/qwen3-5-35b-a3b) | Alibaba | 88.5 | 2 | 21 total |
| 30 | [Gemma 4 26B A4B](/models/gemma-4-26b-a4b) | Google | 88.4 | 1 | 13 total |
| 31 | [GLM-5](/models/glm-5) | Z.AI | 88.3 | 2 | 41 total |
| 32 | [Qwen3.8-Flash-Next](/models/qwen3-8-flash-next) | Alibaba | 88 | 1 | 30 total |
| 33 | [o3](/models/o3) | OpenAI | 87.1 | 1 | 10 total |
| 34 | [Gemini 3.5 Flash](/models/gemini-3-5-flash) | Google | 86.9 | 2 | 39 total |
| 35 | [GPT-5.6 Terra](/models/gpt-5-6-terra) | OpenAI | 86.8 | 1 | 44 total |
| 36 | [dots3-note Preview](/models/dots3-note-preview) | Dots Studio | 86.3 | 2 | 30 total |
| 37 | [GPT-5 (medium)](/models/gpt-5-medium) | OpenAI | 86.1 | 1 | 8 total |
| 38 | [Gemini 3 Pro](/models/gemini-3-pro) | Google | 85.8 | 1 | 19 total |
| 39 | [o1](/models/o1) | OpenAI | 85.7 | 2 | 11 total |
| 40 | [Kimi K2.5](/models/kimi-k2-5) | Moonshot AI | 85.6 | 2 | 49 total |
| 41 | [Qwen3.6 Plus](/models/qwen3-6-plus) | Alibaba | 85.5 | 3 | 53 total |
| 42 | [GPT-5.1-Codex](/models/gpt-5-1-codex) | OpenAI | 85.3 | 1 | 10 total |
| 43 | [Inkling](/models/inkling) | Thinking Machines Lab | 85.2 | 1 | 31 total |
| 44 | [Qwen3.8-27B](/models/qwen3-8-27b) | Alibaba | 84.7 | 1 | 41 total |
| 45 | [Granite 4.2 8B](/models/granite-4-2-8b) | IBM | 84.4 | 1 | 20 total |
| 46 | [GPT-OSS 120B](/models/gpt-oss-120b) | OpenAI | 84 | 1 | 10 total |
| 47 | [MiMo-V2-Pro](/models/mimo-v2-pro) | Xiaomi | 83.7 | 1 | 8 total |
| 48 | [Mistral Medium 3.5 128B](/models/mistral-medium-3-5-128b) | Mistral | 83.7 | 1 | 18 total |
| 49 | [GLM-4.7](/models/glm-4-7) | Z.AI | 82.6 | 1 | 17 total |
| 50 | [Qwen3.6-27B](/models/qwen3-6-27b) | Alibaba | 82.2 | 1 | 46 total |
| 51 | [Step 3.7 Flash](/models/step-3-7-flash) | StepFun | 81.8 | 1 | 19 total |
| 52 | [Muse Glimmer 30B](/models/muse-glimmer-30b) | Meta | 80.1 | 1 | 23 total |
| 53 | [GPT-OSS 20B](/models/gpt-oss-20b) | OpenAI | 78.9 | 1 | 9 total |
| 54 | [Qwen3.5 397B](/models/qwen3-5-397b) | Alibaba | 78.8 | 2 | 35 total |
| 55 | [K-Exaone](/models/k-exaone) | LG AI Research | 78.4 | 1 | 8 total |
| 56 | [Claude Fable 5](/models/claude-fable) | Anthropic | 78.3 | 1 | 30 total |
| 57 | [Qwen3.6-35B-A3B](/models/qwen3-6-35b-a3b) | Alibaba | 78 | 1 | 49 total |
| 58 | [Kimi K2.7 Code](/models/kimi-k2-7-code) | Moonshot AI | 76.3 | 1 | 11 total |
| 59 | [Nemotron 3 Nano Omni 30B A3B](/models/nemotron-3-nano-omni-30b-a3b) | NVIDIA | 75.9 | 2 | 23 total |
| 60 | [Ling 3.0 Flash](/models/ling-3-0-flash) | InclusionAI | 75.6 | 1 | 31 total |
| 61 | [Claude Opus 4.8](/models/claude-opus-4-8) | Anthropic | 75.2 | 1 | 45 total |
| 62 | [GLM-5V-Turbo](/models/glm-5v-turbo) | Z.AI | 73.7 | 1 | 9 total |
| 63 | [Ling 3.0 Flash FP8](/models/ling-3-0-flash-fp8) | InclusionAI | 73.6 | 1 | 4 total |
| 64 | [Nemotron 3.5 Lightning 30B A3B NVFP4](/models/nemotron-3-5-lightning-30b-a3b-nvfp4) | NVIDIA | 72.6 | 1 | 14 total |
| 65 | [Claude Opus 4.5 Thinking](/models/claude-opus-4-5-thinking) | Anthropic | 69.7 | 1 | 9 total |
| 66 | [Trinity-Large-Thinking](/models/trinity-large-thinking) | Arcee AI | 67.5 | 1 | 12 total |
| 67 | [Claude 4.1 Opus Thinking](/models/claude-4-1-opus-thinking) | Anthropic | 66.3 | 1 | 7 total |
| 68 | [Gemini 3 Flash](/models/gemini-3-flash) | Google | 66 | 1 | 18 total |
| 69 | [Grok 4](/models/grok-4) | xAI | 64.1 | 1 | 11 total |
| 70 | [MiMo-V2-Omni](/models/mimo-v2-omni) | Xiaomi | 63.9 | 1 | 9 total |
| 71 | [Grok 4.1 Fast (Reasoning)](/models/grok-4-1-fast-reasoning) | xAI | 62.8 | 1 | 9 total |
| 72 | [Grok 4 Fast (Reasoning)](/models/grok-4-fast-reasoning) | xAI | 60 | 1 | 9 total |
| 73 | [DeepSeek V3.2](/models/deepseek-v3-2) | DeepSeek | 58 | 1 | 13 total |
| 74 | [Gemini 2.5 Pro](/models/gemini-2-5-pro) | Google | 57.7 | 1 | 14 total |
| 75 | [Mistral Small 4](/models/mistral-small-4) | Mistral | 57 | 1 | 9 total |
| 76 | [Hy3 Preview](/models/hy3-preview) | Tencent | 54.8 | 1 | 12 total |
| 77 | [Claude 4 Sonnet](/models/claude-4-sonnet) | Anthropic | 53.4 | 1 | 10 total |
| 78 | [Claude Opus 4.6](/models/claude-opus-4-6) | Anthropic | 52.3 | 1 | 38 total |
| 79 | [Gemma 4 E4B](/models/gemma-4-e4b) | Google | 51.8 | 1 | 9 total |
| 80 | [Qwen3 Max](/models/qwen3-max) | Alibaba | 51.7 | 1 | 8 total |
| 81 | [Llama 4 Maverick](/models/llama-4-maverick) | Meta | 50.3 | 1 | 10 total |
| 82 | [GPT-4.1](/models/gpt-4-1) | OpenAI | 50.3 | 2 | 14 total |
| 83 | [DeepSeek V3.1 (Reasoning)](/models/deepseek-v3-1-reasoning) | DeepSeek | 48.3 | 1 | 7 total |
| 84 | [Kimi K2](/models/kimi-k2) | Moonshot AI | 48.3 | 1 | 9 total |
| 85 | [Grok Code Fast 1](/models/grok-code-fast-1) | xAI | 48.2 | 1 | 7 total |
| 86 | [Claude Sonnet 4.6](/models/claude-sonnet-4-6) | Anthropic | 47.9 | 1 | 31 total |
| 87 | [LFM2.5-2.6B](/models/lfm2-5-2-6b) | LiquidAI | 47.6 | 1 | 13 total |
| 88 | [Ling 2.6 Flash](/models/ling-2-6-flash) | InclusionAI | 46.6 | 2 | 11 total |
| 89 | [MiMo-V2-Flash](/models/mimo-v2-flash) | Xiaomi | 46.2 | 1 | 9 total |
| 90 | [DeepSeek-R1](/models/deepseek-r1) | DeepSeek | 45.9 | 1 | 7 total |
| 91 | [Llama 4 Scout](/models/llama-4-scout) | Meta | 45.7 | 1 | 10 total |
| 92 | [Mistral Medium 3](/models/mistral-medium-3) | Mistral | 45.5 | 1 | 8 total |
| 93 | [Gemini 2.5 Flash](/models/gemini-2-5-flash) | Google | 45.1 | 1 | 10 total |
| 94 | [LFM2.5-8B-A1B](/models/lfm2-5-8b-a1b) | LiquidAI | 44.7 | 3 | 13 total |
| 95 | [GPT-4.1 mini](/models/gpt-4-1-mini) | OpenAI | 44.2 | 2 | 14 total |
| 96 | [Nova Pro](/models/nova-pro) | Amazon | 43.9 | 1 | 8 total |
| 97 | [Gemma 4 E2B](/models/gemma-4-e2b) | Google | 43.8 | 1 | 9 total |
| 98 | [DeepSeek V3.1](/models/deepseek-v3-1) | DeepSeek | 43.5 | 1 | 7 total |
| 99 | [GLM-4.5-Air](/models/glm-4-5-air) | Z.AI | 43.3 | 1 | 7 total |
| 100 | [GLM-4.6](/models/glm-4-6) | Z.AI | 42.1 | 1 | 10 total |
| 101 | [Grok 4.1 Fast](/models/grok-4-1-fast) | xAI | 41.8 | 1 | 8 total |
| 102 | [Mistral Large 3](/models/mistral-large-3) | Mistral | 41.4 | 1 | 9 total |
| 103 | [Claude 3 Haiku](/models/claude-3-haiku) | Anthropic | 41.3 | 1 | 8 total |
| 104 | [DeepSeek V3](/models/deepseek-v3) | DeepSeek | 39.6 | 2 | 9 total |
| 105 | [Claude Opus 4.5](/models/claude-opus-4-5) | Anthropic | 39.5 | 3 | 50 total |
| 106 | [Sarvam 105B](/models/sarvam-105b) | Sarvam | 39.1 | 1 | 7 total |
| 107 | [GPT-4o](/models/gpt-4o) | OpenAI | 39 | 1 | 7 total |
| 108 | [Solar Pro 2](/models/solar-pro-2) | Upstage | 38.2 | 1 | 7 total |
| 109 | [Exaone 4.0 32B](/models/exaone-4-0-32b) | LG AI Research | 37.9 | 1 | 9 total |
| 110 | [GPT-4.1 nano](/models/gpt-4-1-nano) | OpenAI | 36 | 2 | 13 total |
| 111 | [Gemma 3 27B](/models/gemma-3-27b) | Google | 35.7 | 1 | 9 total |
| 112 | [ZAYA1-8B](/models/zaya1-8b) | Zyphra | 35.6 | 2 | 10 total |
| 113 | [Mistral Large 2](/models/mistral-large-2) | Mistral | 35 | 1 | 7 total |
| 114 | [GPT-4o mini](/models/gpt-4o-mini) | OpenAI | 34.7 | 1 | 6 total |
| 115 | [Sarvam 30B](/models/sarvam-30b) | Sarvam | 28.9 | 1 | 7 total |
| 116 | [DeepSeek R1 Distill Qwen 32B](/models/deepseek-r1-distill-qwen-32b) | DeepSeek | 28.5 | 1 | 5 total |
| 117 | [Phi-4](/models/phi-4) | Microsoft | 28.5 | 1 | 7 total |
| 118 | [Granite-4.0-1B](/models/granite-4-0-1b) | IBM | 28.5 | 1 | 7 total |
| 119 | [Granite-4.0-H-1B](/models/granite-4-0-h-1b) | IBM | 28.5 | 1 | 7 total |
| 120 | [Granite-4.0-350M](/models/granite-4-0-350m) | IBM | 28.5 | 1 | 7 total |
| 121 | [Granite-4.0-H-350M](/models/granite-4-0-h-350m) | IBM | 28.5 | 1 | 7 total |
| 122 | [Exaone 4.0 1.2B](/models/exaone-4-0-1-2b) | LG AI Research | 28.5 | 1 | 7 total |
| 123 | [MiniCPM5-1B](/models/minicpm5-1b) | OpenBMB | 24.8 | 2 | 13 total |
| 124 | [LFM2.5-230M](/models/lfm2-5-230m) | LiquidAI | 9.7 | 2 | 6 total |

## Decision-ready shortlist

- #1 [MAI-Thinking-1](/models/mai-thinking-1) — 94.7 weighted score, Proprietary, 256K context.
- #2 [Grok 4.3](/models/grok-4-3) — 94.4 weighted score, Proprietary, 1M context.
- #3 [GPT-5.2-Codex](/models/gpt-5-2-codex) — 93.5 weighted score, Proprietary, 400K context.
- #4 [MiMo-V2.5-Pro](/models/mimo-v2-5-pro) — 93.5 weighted score, Proprietary, 1M context.
- #5 [MiniMax M3](/models/minimax-m3) — 93.5 weighted score, Open Weight, 1M context.

## Benchmarks in this category

### [IFEval](/benchmarks/ifeval) (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.

- Ranking status: Display only
- Year: 2023
- Format: Constrained generation
- Difficulty: Instruction precision

### [IFBench](/benchmarks/ifbench) (Instruction Following Benchmark)

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.

- Ranking status: Weighted (70% of this category)
- Year: 2025
- Format: undefined
- Difficulty: undefined

### [SOB Value Acc](/benchmarks/sobvalueacc) (Structured Output Benchmark Value Accuracy)

A structured-output benchmark from Interfaze measuring whether extracted JSON leaf values exactly match verified ground truth.

- Ranking status: Display only
- Year: 2026
- Format: Value accuracy
- Difficulty: Production structured-output reliability