Instruction Following · Benchmark profile
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.
Data verified 27 Jul 2026 · Methodology 1.8.0
Benchmark score on Instruction-Following Eval
Qwen3.5 27B leads at 95%, followed by Agents-A1 (94.8%) and Qwen3.7-Plus (94.6%).
10 ranked models · higher is better · labels show rank and score
| Model | Rank | Provider | Score |
|---|---|---|---|
| Qwen3.5 27B | #1 | Alibaba Cloud | 95% |
| Agents-A1 | #2 | InternScience | 94.8% |
| Qwen3.7-Plus | #3 | Alibaba Cloud | 94.6% |
| Qwen3.6 Plus | #4 | Alibaba Cloud | 94.3% |
| Qwen3.7-Max | #5 | Alibaba Cloud | 94.3% |
| Kimi K2.5 | #6 | Moonshot AI | 93.9% |
| o3-mini | #7 | OpenAI | 93.9% |
| Qwen3.5-122B-A10B | #8 | Alibaba Cloud | 93.4% |
| GLM-5 | #9 | Z.AI | 92.6% |
| Qwen3.5 397B | #10 | Alibaba Cloud | 92.6% |
One best score per model · higher is better
| Rank | Model | Provider | License | Evidence use | Score |
|---|---|---|---|---|---|
| #1 | Qwen3.5 27B qwen3.5-27b | Alibaba Cloud | open | Reference only | 95% |
| #2 | Agents-A1 agents-a1 | InternScience | open | Estimated reference | 94.8% |
| #3 | Qwen3.7-Plus qwen3.7-plus | Alibaba Cloud | closed | Reference only | 94.6% |
| #4 | Qwen3.6 Plus qwen3-6-plus | Alibaba Cloud | closed | Reference only | 94.3% |
| #5 | Qwen3.7-Max qwen3.7-max | Alibaba Cloud | closed | Reference only | 94.3% |
| #6 | Kimi K2.5 kimi-k2.5 | Moonshot AI | open | Reference only | 93.9% |
| #7 | o3-mini o3-mini | OpenAI | closed | Reference only | 93.9% |
| #8 | Qwen3.5-122B-A10B qwen3-5-122b-a10b | Alibaba Cloud | open | Reference only | 93.4% |
| #9 | GLM-5 glm-5 | Z.AI | open | Reference only | 92.6% |
| #10 | Qwen3.5 397B qwen3.5-397b-a17b | Alibaba Cloud | open | Estimated reference | 92.6% |
| #11 | o1 o1 | OpenAI | closed | Estimated reference | 92.2% |
| #12 | Qwen3.5-35B-A3B qwen3-5-35b-a3b | Alibaba Cloud | open | Reference only | 91.9% |
| #13 | LFM2.5-8B-A1B lfm2-5-8b-a1b | LiquidAI | open | Estimated reference | 91.8% |
| #14 | Claude Opus 4.5 claude-opus-4-5 | Anthropic | closed | Reference only | 90.9% |
| #15 | GPT-4.1 mini gpt-4-1-mini | OpenAI | closed | Estimated reference | 88.5% |
| #16 | GPT-4.1 gpt-4-1 | OpenAI | closed | Reference only | 87.4% |
| #17 | DeepSeek V3 deepseek-v3 | DeepSeek | open | Reference only | 86.1% |
| #18 | ZAYA1-8B zaya1-8b | Zyphra | open | Estimated reference | 85.6% |
| #19 | GPT-4.1 nano gpt-4-1-nano | OpenAI | closed | Estimated reference | 83.2% |
| #20 | MiniCPM5-1B minicpm5-1b | OpenBMB | open | Estimated reference | 80.4% |
| #21 | Mellum2-12B-A2.5B-Thinking mellum2-12b-a2-5b-thinking | JetBrains | open | Estimated reference | 76.5% |
| #22 | Mellum2-12B-A2.5B-Instruct mellum2-12b-a2-5b-instruct | JetBrains | open | Estimated reference | 75.8% |
| #23 | LFM2.5-230M lfm2-5-230m | LiquidAI | open | Estimated reference | 71.7% |
| #24 | LFM2.5-VL-450M lfm2-5-vl-450m | LiquidAI | open | Estimated reference | 61.2% |
The top of this snapshot is led by Qwen3.5 27B at 95%; third place is 0.40 points behind. The top-10 spread is 2.4 points.
About 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. Results stay tied to the exact model variant and evaluation system. Multiple systems for the same model use the best published score on this page; overall Lumina scoring uses the median of ranking-eligible rows.
Open benchmark source ↗FAQ
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.
Qwen3.5 27B by Alibaba Cloud currently leads with 95%.
24 models in the LuminaBench cohort have a qualifying score on this benchmark.
No. This benchmark is display-only and does not enter the overall Lumina composite.
Related
