Reasoning
ForecastBench Leaderboard
ForecastBench, from the Forecasting Research Institute, asks models to predict real-world future events, with questions that only resolve after the model's training cutoff so the answers can't have been memorised. It is scored as a difficulty-adjusted Brier index, where higher is better.
Source: epoch24 open models ranked+71 proprietaryData through Sep 2026
Open models ranked on ForecastBench
# shows rank among open models / rank overall (including proprietary).
| # | Model | Score | VRAM @ Q4 |
|---|---|---|---|
| 1 / 29 | Kimi K3 · 2779.9B | 64.2 | ~1674 GB |
| 2 / 35 | MiniMax M3 · 427.0B | 64.0 | ~257 GB |
| 3 / 40 | Qwen3 235B A22B · 235.1B | 63.9 | ~142 GB |
| 4 / 44 | GLM 5 · 753.9B | 63.4 | ~457 GB |
| 5 / 59 | DeepSeek v3 · 684.5B | 62.4 | ~415 GB |
| 6 / 61 | Kimi K2 Instruct 0905 · 1026.4B | 62.4 | ~620 GB |
| 7 / 62 | Llama 3.1 405B Instruct · 405.9B | 62.3 | ~268 GB |
| 8 / 66 | Kimi K2 Instruct · 1026.4B | 62.0 | ~620 GB |
| 9 / 67 | Qwen3 235B A22B Thinking 2507 · 235.1B | 62.0 | ~142 GB |
| 10 / 68 | GLM 4.5 Air · 110.5B | 61.7 | ~67 GB |
| 11 / 70 | Mistral Large Instruct 2407 · 122.6B | 61.5 | ~81 GB |
| 12 / 71 | Qwen2.5 72B Instruct · 72.7B | 61.4 | ~45 GB |
| 13 / 72 | Mistral Large Instruct 2411 · 122.6B | 61.3 | ~75 GB |
| 14 / 74 | Gemma 4 31B IT · 31.3B | 61.1 | ~20 GB |
| 15 / 76 | Llama 3.3 70B Instruct · 70.6B | 61.0 | ~47 GB |
| 16 / 78 | DeepSeek V4 Pro · 1598.8B | 60.7 | ~960 GB |
| 17 / 79 | DeepSeek V3.1 · 684.5B | 60.5 | ~415 GB |
| 18 / 81 | Meta Llama 3 70B · 70.6B | 59.9 | ~47 GB |
| 19 / 82 | Qwen1.5 110B Chat · 111.2B | 59.3 | ~68 GB |
| 20 / 83 | Meta Llama 3 8B Instruct · 8.0B | 59.1 | ~5.3 GB |
| 21 / 84 | Llama 4 Scout 17B 16E Instruct · 108.6B | 58.9 | ~72 GB |
| 22 / 88 | Mixtral 8x22B Instruct v0.1 · 140.6B | 58.6 | ~85 GB |
| 23 / 91 | Mixtral 8x7B Instruct v0.1 · 46.7B | 58.0 | ~29 GB |
| 24 / 94 | Llama 2 70B Chat HF · 69.0B | 50.2 | ~152 GB |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Meta Llama 3 8B Instruct, 8B, score 59.1 — on the efficiency frontier (best score at its size or smaller).
- Gemma 4 31B IT, 31B, score 61.1 — on the efficiency frontier (best score at its size or smaller).
- Qwen2.5 72B Instruct, 73B, score 61.4 — on the efficiency frontier (best score at its size or smaller).
- GLM 4.5 Air, 110B, score 61.7 — on the efficiency frontier (best score at its size or smaller).
- Qwen3 235B A22B, 235B, score 63.9 — on the efficiency frontier (best score at its size or smaller).
- MiniMax M3, 427B, score 64.0 — on the efficiency frontier (best score at its size or smaller).
- Kimi K3, 2.8T, score 64.2 — on the efficiency frontier (best score at its size or smaller).
ForecastBench: frequently asked questions
- What is the best open LLM on ForecastBench?
- Kimi K3 is the top open model on ForecastBench, scoring 64.2. Among all models tested — including proprietary ones — it ranks #27. The top model overall is GPT 5.5 (high) (OpenAI) at 66.7.
- What's the best ForecastBench model you can run on a 24 GB GPU?
- Gemma 4 31B IT is the highest-scoring open model that fits in 24 GB at 4-bit quantization (~20 GB), scoring 61.1 on ForecastBench.
- What's the best ForecastBench model you can run on a 12 GB GPU?
- Meta Llama 3 8B Instruct is the highest-scoring open model that fits in 12 GB at 4-bit quantization (~5.3 GB), scoring 59.1 on ForecastBench.
- Can open models match proprietary models on ForecastBench?
- Not quite on ForecastBench: the strongest proprietary model (GPT 5.5 (high)) scores 66.7, ahead of the best open model (Kimi K3) at 64.2 — 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.