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: epoch67 open models ranked+143 proprietaryData through Sep 2026

Open models ranked on DTBench

# shows rank among open models / rank overall (including proprietary).

#ModelScore
1 / 44DeepSeek V4 Pro 0813 · 1650.5B
93.9%
2 / 46GLM 5.2 · 753.3B
93.6%
3 / 57Kimi K3 · 2779.9B
91.2%
4 / 58DeepSeek V4 Flash 0731 · 304.2B
90.9%
5 / 60Kimi K2.6 · 1026.9B
90.9%
6 / 61DeepSeek V4 Pro · 1598.8B
90.7%
7 / 68NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 560.5B
90.1%
8 / 71DeepSeek V4.1 Flash · 763.2B
89.9%
9 / 76Qwen3.8 27B · 27.8B
88.0%
10 / 77DeepSeek V3.2 Exp · 685.4B
87.7%
11 / 79GLM 5.3 · 753.3B
87.7%
12 / 80Inkling · 952.4B
87.5%
13 / 81Qwen3.5 397B A17B · 403.4B
87.5%
14 / 85DeepSeek V4 Flash · 290.9B
86.4%
15 / 86DeepSeek V3.2 · 685.4B
85.6%
16 / 89MiMo V2.5 Pro · 1023.2B
84.5%
17 / 90Qwen3.5 122B A10B · 125.1B
84.3%
18 / 97DeepSeek V3.1 · 684.5B
82.7%
19 / 98Gemma 4 31B IT · 31.3B
82.7%
20 / 101Qwen3.5 27B · 27.8B
82.4%
21 / 105DeepSeek V3.1 Terminus · 684.5B
81.3%
22 / 110Qwen3 235B A22B Thinking 2507 · 235.1B
80.3%
23 / 114Qwen3.5 35B A3B · 36.0B
80.0%
24 / 116MiniMax M3 · 427.0B
78.9%
25 / 118Qwen3 235B A22B Instruct 2507 · 235.1B
78.4%
26 / 119Qwen3.6 27B · 27.8B
78.1%
27 / 125GPT OSS 120B · 116.8B
76.3%
28 / 126Qwen3 235B A22B · 235.1B
75.7%
29 / 128Gemma 4 26B A4B IT · 25.8B
74.9%
30 / 130Qwen3.6 35B A3B · 36.0B
73.9%
31 / 133Qwen3.5 9B · 9.7B
71.2%
32 / 135Qwen3 30B A3B Thinking 2507 · 30.5B
69.3%
33 / 140GPT OSS 20B · 20.9B
68.0%
34 / 144Qwen3 32B · 32.8B
67.5%
35 / 145Qwen3 30B A3B Instruct 2507 · 30.5B
67.2%
36 / 147Mistral Large 3 675B Instruct 2512 · 675B
65.1%
37 / 148DeepSeek v3 0324 · 684.5B
64.8%
38 / 150Qwen3 14B · 14.8B
64.0%
39 / 152Qwen2.5 72B Instruct · 72.7B
62.9%
40 / 158Llama 4 Maverick 17B 128E Instruct · 401.6B
61.9%
41 / 161Llama 3.1 405B Instruct · 405.9B
61.4%
42 / 162C4ai Command A 03 2025 · 111.1B
61.3%
43 / 163Magistral Small 2509 · 24.0B
61.3%
44 / 164Mistral Large Instruct 2407 · 122.6B
61.2%
45 / 165Mistral Large Instruct 2411 · 122.6B
60.8%
46 / 166Qwen3 30B A3B · 30.5B
60.3%
47 / 167Llama 3.1 70B Instruct · 70.6B
60.0%
48 / 168Mistral Small 3.2 24B Instruct 2506 · 24.0B
59.9%
49 / 169Qwen3 8B · 8.2B
59.7%
50 / 170Llama 3.3 70B Instruct · 70.6B
59.5%
51 / 172Mistral Small 3.1 24B Instruct 2503 · 24.0B
58.6%
52 / 173Llama 4 Scout 17B 16E Instruct · 108.6B
57.9%
53 / 177Mixtral 8x22B Instruct v0.1 · 140.6B
55.1%
54 / 178C4ai Command R Plus 08 2024 · 103.8B
54.9%
55 / 180Meta Llama 3 70B Instruct · 70.6B
54.2%
56 / 186Gemma 3 27B IT · 27.4B
52.5%
57 / 192Gemma 3 4B IT · 4.3B
50.9%
58 / 193Llama 3.1 8B Instruct · 8.0B
50.9%
59 / 196Mixtral 8x7B Instruct v0.1 · 46.7B
49.6%
60 / 197Mistral Small Instruct 2409 · 22.2B
49.5%
61 / 198Gemma 3 12B IT · 12.2B
48.8%
62 / 199Mistral Nemo Instruct 2407 · 12.2B
48.6%
63 / 202Qwen2.5 7B Instruct · 7.6B
47.7%
64 / 203C4ai Command R 08 2024 · 32.3B
46.4%
65 / 207Meta Llama 3 8B Instruct · 8.0B
43.9%
66 / 209Llama 2 13B Chat HF · 13.0B
42.2%
67 / 210Llama 2 70B Chat HF · 69.0B
41.6%

Score vs model size

Which models give the most quality for their size — the ones worth running locally.

10B100B1Tmodel size (log scale) →93.9%41.6%Kimi K3 · 2.8T · 91.2%Kimi K2.6 · 1T · 90.9%DeepSeek V4 Pro · 1.6T · 90.7%NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 561B · 90.1%DeepSeek V4.1 Flash · 763B · 89.9%DeepSeek V3.2 Exp · 685B · 87.7%GLM 5.3 · 753B · 87.7%Qwen3.5 397B A17B · 403B · 87.5%Inkling · 952B · 87.5%DeepSeek V4 Flash · 291B · 86.4%DeepSeek V3.2 · 685B · 85.6%MiMo V2.5 Pro · 1T · 84.5%Qwen3.5 122B A10B · 125B · 84.3%DeepSeek V3.1 · 685B · 82.7%Gemma 4 31B IT · 31B · 82.7%Qwen3.5 27B · 28B · 82.4%DeepSeek V3.1 Terminus · 685B · 81.3%Qwen3 235B A22B Thinking 2507 · 235B · 80.3%Qwen3.5 35B A3B · 36B · 80.0%MiniMax M3 · 427B · 78.9%Qwen3 235B A22B Instruct 2507 · 235B · 78.4%Qwen3.6 27B · 28B · 78.1%GPT OSS 120B · 117B · 76.3%Qwen3 235B A22B · 235B · 75.7%Qwen3.6 35B A3B · 36B · 73.9%Qwen3 30B A3B Thinking 2507 · 31B · 69.3%GPT OSS 20B · 21B · 68.0%Qwen3 32B · 33B · 67.5%Qwen3 30B A3B Instruct 2507 · 31B · 67.2%Mistral Large 3 675B Instruct 2512 · 675B · 65.1%DeepSeek v3 0324 · 685B · 64.8%Qwen3 14B · 15B · 64.0%Qwen2.5 72B Instruct · 73B · 62.9%Llama 4 Maverick 17B 128E Instruct · 402B · 61.9%Llama 3.1 405B Instruct · 406B · 61.4%C4ai Command A 03 2025 · 111B · 61.3%Magistral Small 2509 · 24B · 61.3%Mistral Large Instruct 2407 · 123B · 61.2%Mistral Large Instruct 2411 · 123B · 60.8%Qwen3 30B A3B · 31B · 60.3%Llama 3.1 70B Instruct · 71B · 60.0%Mistral Small 3.2 24B Instruct 2506 · 24B · 59.9%Llama 3.3 70B Instruct · 71B · 59.5%Mistral Small 3.1 24B Instruct 2503 · 24B · 58.6%Llama 4 Scout 17B 16E Instruct · 109B · 57.9%Mixtral 8x22B Instruct v0.1 · 141B · 55.1%C4ai Command R Plus 08 2024 · 104B · 54.9%Meta Llama 3 70B Instruct · 71B · 54.2%Gemma 3 27B IT · 27B · 52.5%Llama 3.1 8B Instruct · 8B · 50.9%Mixtral 8x7B Instruct v0.1 · 47B · 49.6%Mistral Small Instruct 2409 · 22B · 49.5%Gemma 3 12B IT · 12B · 48.8%Mistral Nemo Instruct 2407 · 12B · 48.6%Qwen2.5 7B Instruct · 8B · 47.7%C4ai Command R 08 2024 · 32B · 46.4%Meta Llama 3 8B Instruct · 8B · 43.9%Llama 2 13B Chat HF · 13B · 42.2%Llama 2 70B Chat HF · 69B · 41.6%Gemma 3 4B IT · 4B · 50.9%Gemma 3 4B ITQwen3 8B · 8B · 59.7%Qwen3 8BQwen3.5 9B · 10B · 71.2%Qwen3.5 9BGemma 4 26B A4B IT · 26B · 74.9%Gemma 4 26B A4B ITQwen3.8 27B · 28B · 88.0%Qwen3.8 27BDeepSeek V4 Flash 0731 · 304B · 90.9%DeepSeek V4 Flash 0731GLM 5.2 · 753B · 93.6%GLM 5.2DeepSeek V4 Pro 0813 · 1.7T · 93.9%DeepSeek V4 Pro 0813
Each dot is a model. Up = higher score, left = smaller (easier to run locally). The dashed line marks the efficiency frontier — the best score you can get at each size or smaller.
  • 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 #44. The top model overall is Claude Fable 5 Max (Anthropic) at 98.4%.
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 (about 15 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 (about 5 GB), scoring 71.2% on DTBench.
Can open models match proprietary models on DTBench?
Not quite on DTBench: the strongest proprietary model (Claude Fable 5 Max) scores 98.4%, 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.