Reasoning
DTBench Leaderboard
DTBench is a set of handcrafted, expert-validated multiple-choice questions on decision theory — Newcomb-like problems and related puzzles — built by the Conceptual Reasoning Index project with Anthropic. It tests whether a model can follow an abstract argument rather than recall a fact.
Source: epoch70 open models ranked+165 proprietaryData through Oct 2026
Open models ranked on DTBench
# shows rank among open models / rank overall (including proprietary).
| # | Model | Score | VRAM @ Q4 |
|---|---|---|---|
| 1 / 61 | DeepSeek V4 Pro 0813 · 1650.5B | 93.9% | ~991 GB |
| 2 / 63 | GLM 5.2 · 753.3B | 93.6% | ~456 GB |
| 3 / 75 | Kimi K3 · 2779.9B | 91.2% | ~1674 GB |
| 4 / 76 | DeepSeek V4 Flash 0731 · 304.2B | 90.9% | ~183 GB |
| 5 / 78 | Kimi K2.6 · 1026.9B | 90.9% | ~620 GB |
| 6 / 79 | DeepSeek V4 Pro · 1598.8B | 90.7% | ~960 GB |
| 7 / 87 | NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 560.5B | 90.1% | ~370 GB |
| 8 / 90 | DeepSeek V4.1 Flash · 763.2B | 89.9% | ~458 GB |
| 9 / 95 | Mistral Large 4 1T A52B · 1000B | 88.2% | ~604 GB |
| 10 / 96 | Qwen3.8 27B · 27.8B | 88.0% | ~17 GB |
| 11 / 97 | DeepSeek V3.2 Exp · 685.4B | 87.7% | ~415 GB |
| 12 / 99 | GLM 5.3 · 753.3B | 87.7% | ~456 GB |
| 13 / 100 | Inkling · 952.4B | 87.5% | ~572 GB |
| 14 / 101 | Qwen3.5 397B A17B · 403.4B | 87.5% | ~242 GB |
| 15 / 107 | DeepSeek V4 Flash · 290.9B | 86.4% | ~175 GB |
| 16 / 108 | DeepSeek V3.2 · 685.4B | 85.6% | ~415 GB |
| 17 / 111 | MiMo V2.5 Pro · 1023.2B | 84.5% | ~614 GB |
| 18 / 113 | Qwen3.5 122B A10B · 125.1B | 84.3% | ~75 GB |
| 19 / 120 | DeepSeek V3.1 · 684.5B | 82.7% | ~415 GB |
| 20 / 121 | Gemma 4 31B IT · 31.3B | 82.7% | ~20 GB |
| 21 / 124 | Qwen3.5 27B · 27.8B | 82.4% | ~17 GB |
| 22 / 128 | DeepSeek V3.1 Terminus · 684.5B | 81.3% | ~415 GB |
| 23 / 134 | Qwen3 235B A22B Thinking 2507 · 235.1B | 80.3% | ~142 GB |
| 24 / 138 | Qwen3.5 35B A3B · 36.0B | 80.0% | ~22 GB |
| 25 / 140 | MiniMax M3 · 427.0B | 78.9% | ~257 GB |
| 26 / 142 | Qwen3 235B A22B Instruct 2507 · 235.1B | 78.4% | ~142 GB |
| 27 / 143 | Qwen3.6 27B · 27.8B | 78.1% | ~17 GB |
| 28 / 149 | GPT OSS 120B · 116.8B | 76.3% | ~71 GB |
| 29 / 150 | Qwen3 235B A22B · 235.1B | 75.7% | ~142 GB |
| 30 / 151 | Mistral Medium 3.5 128B · 127.7B | 75.5% | ~78 GB |
| 31 / 153 | Gemma 4 26B A4B IT · 25.8B | 74.9% | ~16 GB |
| 32 / 155 | Qwen3.6 35B A3B · 36.0B | 73.9% | ~22 GB |
| 33 / 158 | Qwen3.5 9B · 9.7B | 71.2% | ~6.4 GB |
| 34 / 159 | Mistral Small 4 119B 2603 · 119.4B | 70.9% | ~73 GB |
| 35 / 160 | Qwen3 30B A3B Thinking 2507 · 30.5B | 69.3% | ~19 GB |
| 36 / 165 | GPT OSS 20B · 20.9B | 68.0% | ~13 GB |
| 37 / 169 | Qwen3 32B · 32.8B | 67.5% | ~20 GB |
| 38 / 170 | Qwen3 30B A3B Instruct 2507 · 30.5B | 67.2% | ~19 GB |
| 39 / 172 | Mistral Large 3 675B Instruct 2512 · 675B | 65.1% | ~446 GB |
| 40 / 173 | DeepSeek v3 0324 · 684.5B | 64.8% | ~415 GB |
| 41 / 175 | Qwen3 14B · 14.8B | 64.0% | ~9.5 GB |
| 42 / 177 | Qwen2.5 72B Instruct · 72.7B | 62.9% | ~45 GB |
| 43 / 183 | Llama 4 Maverick 17B 128E Instruct · 401.6B | 61.9% | ~265 GB |
| 44 / 186 | Llama 3.1 405B Instruct · 405.9B | 61.4% | ~268 GB |
| 45 / 187 | C4ai Command A 03 2025 · 111.1B | 61.3% | ~73 GB |
| 46 / 188 | Magistral Small 2509 · 24.0B | 61.3% | ~15 GB |
| 47 / 189 | Mistral Large Instruct 2407 · 122.6B | 61.2% | ~81 GB |
| 48 / 190 | Mistral Large Instruct 2411 · 122.6B | 60.8% | ~75 GB |
| 49 / 191 | Qwen3 30B A3B · 30.5B | 60.3% | ~19 GB |
| 50 / 192 | Llama 3.1 70B Instruct · 70.6B | 60.0% | ~47 GB |
| 51 / 193 | Mistral Small 3.2 24B Instruct 2506 · 24.0B | 59.9% | ~15 GB |
| 52 / 194 | Qwen3 8B · 8.2B | 59.7% | ~5.5 GB |
| 53 / 195 | Llama 3.3 70B Instruct · 70.6B | 59.5% | ~47 GB |
| 54 / 197 | Mistral Small 3.1 24B Instruct 2503 · 24.0B | 58.6% | ~15 GB |
| 55 / 198 | Llama 4 Scout 17B 16E Instruct · 108.6B | 57.9% | ~72 GB |
| 56 / 202 | Mixtral 8x22B Instruct v0.1 · 140.6B | 55.1% | ~85 GB |
| 57 / 203 | C4ai Command R Plus 08 2024 · 103.8B | 54.9% | ~69 GB |
| 58 / 205 | Meta Llama 3 70B Instruct · 70.6B | 54.2% | ~47 GB |
| 59 / 211 | Gemma 3 27B IT · 27.4B | 52.5% | ~18 GB |
| 60 / 217 | Gemma 3 4B IT · 4.3B | 50.9% | ~2.8 GB |
| 61 / 218 | Llama 3.1 8B Instruct · 8.0B | 50.9% | ~5.3 GB |
| 62 / 221 | Mixtral 8x7B Instruct v0.1 · 46.7B | 49.6% | ~29 GB |
| 63 / 222 | Mistral Small Instruct 2409 · 22.2B | 49.5% | ~14 GB |
| 64 / 223 | Gemma 3 12B IT · 12.2B | 48.8% | ~8.0 GB |
| 65 / 224 | Mistral Nemo Instruct 2407 · 12.2B | 48.6% | ~8.1 GB |
| 66 / 227 | Qwen2.5 7B Instruct · 7.6B | 47.7% | ~5.0 GB |
| 67 / 228 | C4ai Command R 08 2024 · 32.3B | 46.4% | ~21 GB |
| 68 / 232 | Meta Llama 3 8B Instruct · 8.0B | 43.9% | ~5.3 GB |
| 69 / 234 | Llama 2 13B Chat HF · 13.0B | 42.2% | ~8.6 GB |
| 70 / 235 | Llama 2 70B Chat HF · 69.0B | 41.6% | ~152 GB |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Gemma 3 4B IT, 4B, score 50.9% — on the efficiency frontier (best score at its size or smaller).
- Qwen3 8B, 8B, score 59.7% — on the efficiency frontier (best score at its size or smaller).
- Qwen3.5 9B, 10B, score 71.2% — on the efficiency frontier (best score at its size or smaller).
- Gemma 4 26B A4B IT, 26B, score 74.9% — on the efficiency frontier (best score at its size or smaller).
- Qwen3.8 27B, 28B, score 88.0% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek V4 Flash 0731, 304B, score 90.9% — on the efficiency frontier (best score at its size or smaller).
- GLM 5.2, 753B, score 93.6% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek V4 Pro 0813, 1.7T, score 93.9% — on the efficiency frontier (best score at its size or smaller).
DTBench: frequently asked questions
- What is the best open LLM on DTBench?
- DeepSeek V4 Pro 0813 is the top open model on DTBench, scoring 93.9%. Among all models tested — including proprietary ones — it ranks #60. The top model overall is Claude Opus 5.5 Max (Anthropic) at 98.9%.
- What's the best DTBench 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 88.0% on DTBench.
- What's the best DTBench model you can run on a 12 GB GPU?
- Qwen3.5 9B is the highest-scoring open model that fits in 12 GB at 4-bit quantization (~6.4 GB), scoring 71.2% on DTBench.
- Can open models match proprietary models on DTBench?
- Not quite on DTBench: the strongest proprietary model (Claude Opus 5.5 Max) scores 98.9%, ahead of the best open model (DeepSeek V4 Pro 0813) at 93.9% — 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.