Reasoning
SimpleBench Leaderboard
SimpleBench is a set of everyday, common-sense and trick questions that humans answer easily but language models often get wrong. It probes basic reasoning and robustness rather than specialist knowledge.
Source: epoch25 open models ranked+76 proprietaryData through Aug 2026
Open models ranked on SimpleBench
# shows rank among open models / rank overall (including proprietary).
| # | Model | Score | VRAM @ Q4 |
|---|---|---|---|
| 1 / 27 | DeepSeek V4 Flash 0731 · 304.2B | 61.1% | ~183 GB |
| 2 / 29 | Kimi K3 · 2779.9B | 60.7% | ~1674 GB |
| 3 / 36 | GLM 5.2 · 753.3B | 58.8% | ~456 GB |
| 4 / 37 | Kimi K2.7 Code · 1026.9B | 57.9% | ~620 GB |
| 5 / 42 | GLM 5.1 · 753.9B | 55.1% | ~457 GB |
| 6 / 45 | GLM 5 · 753.9B | 53.2% | ~457 GB |
| 7 / 48 | DeepSeek V3.2 Speciale · 685.4B | 52.6% | ~415 GB |
| 8 / 51 | Inkling · 952.4B | 50.0% | ~572 GB |
| 9 / 54 | GLM 4.7 · 358.3B | 47.7% | ~216 GB |
| 10 / 57 | Kimi K2.5 · 1026.9B | 46.8% | ~620 GB |
| 11 / 60 | DeepSeek V4 Flash · 290.9B | 46.3% | ~175 GB |
| 12 / 63 | MiniMax M3 · 427.0B | 45.8% | ~257 GB |
| 13 / 70 | DeepSeek R1 0528 · 684.5B | 40.8% | ~415 GB |
| 14 / 72 | DeepSeek V3.1 · 684.5B | 40.0% | ~415 GB |
| 15 / 79 | Qwen3 235B A22B · 235.1B | 31.0% | ~142 GB |
| 16 / 80 | DeepSeek R1 · 684.5B | 30.9% | ~415 GB |
| 17 / 82 | Llama 4 Maverick 17B 128E Instruct · 401.6B | 27.7% | ~265 GB |
| 18 / 84 | DeepSeek v3 0324 · 684.5B | 27.2% | ~415 GB |
| 19 / 87 | Kimi K2 Instruct · 1026.4B | 26.3% | ~620 GB |
| 20 / 90 | Llama 3.1 405B Instruct · 405.9B | 23.0% | ~268 GB |
| 21 / 93 | Mistral Large Instruct 2407 · 122.6B | 22.5% | ~81 GB |
| 22 / 94 | GPT OSS 120B · 116.8B | 22.1% | ~71 GB |
| 23 / 95 | Llama 3.3 70B Instruct · 70.6B | 19.9% | ~47 GB |
| 24 / 96 | DeepSeek v3 · 684.5B | 18.9% | ~415 GB |
| 25 / 100 | C4ai Command R Plus 08 2024 · 103.8B | 17.4% | ~69 GB |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Llama 3.3 70B Instruct, 71B, score 19.9% — on the efficiency frontier (best score at its size or smaller).
- GPT OSS 120B, 117B, score 22.1% — on the efficiency frontier (best score at its size or smaller).
- Mistral Large Instruct 2407, 123B, score 22.5% — on the efficiency frontier (best score at its size or smaller).
- Qwen3 235B A22B, 235B, score 31.0% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek V4 Flash, 291B, score 46.3% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek V4 Flash 0731, 304B, score 61.1% — on the efficiency frontier (best score at its size or smaller).
SimpleBench: frequently asked questions
- What is the best open LLM on SimpleBench?
- DeepSeek V4 Flash 0731 is the top open model on SimpleBench, scoring 61.1%. Among all models tested — including proprietary ones — it ranks #27. The top model overall is Claude Fable 5 Max (Anthropic) at 81.9%.
- Can open models match proprietary models on SimpleBench?
- Not quite on SimpleBench: the strongest proprietary model (Claude Fable 5 Max) scores 81.9%, ahead of the best open model (DeepSeek V4 Flash 0731) at 61.1% — 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.