Reasoning
ARC-AGI-2 Leaderboard
ARC-AGI-2 is the second-generation ARC Prize puzzle set, rebuilt so that tasks solved by brute-force search or memorization were removed and every problem still requires inferring a novel rule from a handful of examples. It is markedly harder than the original ARC-AGI, and most models — open and closed alike — score near zero.
Source: epoch22 open models ranked+214 proprietaryData through Oct 2026
Open models ranked on ARC-AGI-2
# shows rank among open models / rank overall (including proprietary).
| # | Model | Score | VRAM @ Q4 |
|---|---|---|---|
| 1 / 42 | DeepSeek V4.1 Flash 1017 · 763.2B | 78.1% | ~458 GB |
| 2 / 51 | DeepSeek V4.1 Flash · 763.2B | 72.9% | ~458 GB |
| 3 / 66 | GLM 5.3 Flash · 321.3B | 65.8% | ~195 GB |
| 4 / 72 | DeepSeek V4 Flash 0731 · 304.2B | 61.4% | ~183 GB |
| 5 / 75 | DeepSeek V4 Pro 0813 · 1650.5B | 61.3% | ~991 GB |
| 6 / 79 | Kimi K3 · 2779.9B | 60.4% | ~1674 GB |
| 7 / 87 | Muse Spark Open 816B A42B · 822.0B | 57.5% | ~543 GB |
| 8 / 98 | Qwen3.8 27B · 27.8B | 42.4% | ~17 GB |
| 9 / 101 | Inkling Small · 266.0B | 40.1% | ~160 GB |
| 10 / 105 | Inkling · 952.4B | 36.5% | ~572 GB |
| 11 / 120 | Beam 1.0 501B A23B · 501.3B | 23.5% | ~301 GB |
| 12 / 122 | GLM 5.2 · 753.3B | 22.8% | ~456 GB |
| 13 / 133 | Kimi K2.5 · 1026.9B | 11.8% | ~620 GB |
| 14 / 156 | GLM 5 · 753.9B | 4.9% | ~457 GB |
| 15 / 157 | MiniMax M2.5 · 228.7B | 4.9% | ~138 GB |
| 16 / 164 | DeepSeek V3.2 · 685.4B | 4.0% | ~415 GB |
| 17 / 196 | DeepSeek R1 · 684.5B | 1.3% | ~415 GB |
| 18 / 203 | Qwen3 235B A22B Instruct 2507 · 235.1B | 1.3% | ~142 GB |
| 19 / 204 | DeepSeek R1 0528 · 684.5B | 1.1% | ~415 GB |
| 20 / 232 | Llama 4 Maverick 17B 128E Instruct · 401.6B | 0.0% | ~265 GB |
| 21 / 233 | Llama 4 Scout 17B 16E Instruct · 108.6B | 0.0% | ~72 GB |
| 22 / 235 | Magistral Small 2506 · 23.6B | 0.0% | ~15 GB |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Magistral Small 2506, 24B, score 0.0% — on the efficiency frontier (best score at its size or smaller).
- Qwen3.8 27B, 28B, score 42.4% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek V4 Flash 0731, 304B, score 61.4% — on the efficiency frontier (best score at its size or smaller).
- GLM 5.3 Flash, 321B, score 65.8% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek V4.1 Flash 1017, 763B, score 78.1% — on the efficiency frontier (best score at its size or smaller).
ARC-AGI-2: frequently asked questions
- What is the best open LLM on ARC-AGI-2?
- DeepSeek V4.1 Flash 1017 is the top open model on ARC-AGI-2, scoring 78.1%. Among all models tested — including proprietary ones — it ranks #42. The top model overall is GPT 6 Astra Max (OpenAI) at 95.0%.
- What's the best ARC-AGI-2 model you can run on a 24 GB GPU?
- Qwen3.8 27B is the highest-scoring open model that fits in 24 GB at 4-bit quantization (~17 GB), scoring 42.4% on ARC-AGI-2.
- Can open models match proprietary models on ARC-AGI-2?
- Not quite on ARC-AGI-2: the strongest proprietary model (GPT 6 Astra Max) scores 95.0%, ahead of the best open model (DeepSeek V4.1 Flash 1017) at 78.1% — but you can run the open one yourself.
Scores aggregated from epoch. llmrun does not run this benchmark — see the source for methodology, or the about benchmarks for what it measures.