Arena Hard
AI benchmark results and model performance
Type: percentage
Recorded results
49
Average score
63.25%
Score range
26.70% - 95.60%
Leading model
95.60% - Qwen 3 235B A22B
Scores over time
Individual benchmark scores plotted by date.
Model results
| Organisation | Models | Reported | Top score | Info | Self reported | Source |
|---|---|---|---|---|---|---|
| Qwen 3 235B A22B | - | 95.60% | - | Yes | Source | |
| Qwen 3 32B | - | 93.80% | - | Yes | Source | |
| Qwen 3 30B A3B | - | 91% | - | Yes | Source | |
| Qwen 3 30B A3B Instruct 2507 | - | 91% | inferred version-family alias from qwen3-30b-a3b | Yes | Source | |
| Qwen 3 30B A3B Thinking 2507 | - | 91% | inferred version-family alias from qwen3-30b-a3b | Yes | Source | |
| Qwen 3 Coder 30B A3B Instruct | - | 91% | inferred high-confidence family alias from qwen3-30b-a3b (score=0.5007; benches=8) | Yes | Source | |
| Qwen 3 Omni 30B A3B Captioner | - | 91% | inferred family alias from qwen3-30b-a3b (score=0.4129; benches=8) | Yes | Source | |
| Qwen 3 Omni 30B A3B Instruct | - | 91% | inferred high-confidence family alias from qwen3-30b-a3b (score=0.4819; benches=8) | Yes | Source | |
| Qwen 3 Omni 30B A3B Thinking | - | 91% | inferred high-confidence family alias from qwen3-30b-a3b (score=0.4819; benches=8) | Yes | Source | |
| Llama 3.3 Nemotron Super 49B v1 | 18 Mar 2025 | 88.30% | - | Yes | Source | |
| Llama 3.3 Nemotron Super 49B V1.5 | - | 88.30% | inferred version-family alias from llama-3.3-nemotron-super-49b-v1 | Yes | Source | |
| Qwen 72B | - | 81.20% | inferred family alias from qwen-2.5-72b-instruct (score=0.3060; benches=14) | Yes | Source | |
| Phi 4 Reasoning Plus | 30 Apr 2025 | 79% | - | Yes | Source | |
| DeepSeek V2.5 (2024-09-05) | 05 Sept 2024 | 76.20% | inferred alias from deepseek-v2.5 | Yes | Source | |
| DeepSeek V2.5 (2024-12-10) | 10 Dec 2024 | 76.20% | inferred alias from deepseek-v2.5 | Yes | Source | |
| Phi 1 | - | 75.40% | inferred family alias from phi-4 (score=0.3100; benches=13) | Yes | Source | |
| Phi 2 | - | 75.40% | inferred family alias from phi-4 (score=0.3100; benches=13) | Yes | Source | |
| Phi 4 | 12 Dec 2024 | 75.40% | - | Yes | Source | |
| Phi 4 Reasoning | 30 Apr 2025 | 73.30% | - | Yes | Source | |
| Granite 4.1 30B | 29 Apr 2026 | 71.02% | - | Yes | Source | |
| Granite 4.1 8B | 29 Apr 2026 | 68.98% | - | Yes | Source | |
| Mistral Small 1.0 | 26 Feb 2024 | 58.30% | inferred family alias from mistral-small-latest (score=0.3650; benches=9) | Yes | Source | |
| Mistral Small 2.0 | 17 Sept 2024 | 58.30% | inferred family alias from mistral-small-latest (score=0.3650; benches=9) | Yes | Source | |
| Mistral Small 4 | 16 Mar 2026 | 58.30% | - | Yes | Source | |
| Mistral Small Creative | 16 Dec 2025 | 58.30% | inferred family alias from mistral-small-latest (score=0.4273; benches=9) | Yes | Source |
