Reasoning
BALROG Leaderboard
BALROG (Benchmarking Agentic LLM and VLM Reasoning On Games) tests long-horizon agentic reasoning across six game environments — BabyAI, Crafter, TextWorld, Baba Is AI, MiniHack and NetHack — scoring the average progress a model makes toward completing each game. It was built by UCL's DARK lab and collaborators.
Source: epoch15 open models ranked+25 proprietaryData through Sep 2026
Open models ranked on BALROG
# shows rank among open models / rank overall (including proprietary).
| # | Model | Score | VRAM @ Q4 |
|---|---|---|---|
| 1 / 14 | DeepSeek R1 · 684.5B | 34.9% | ~415 GB |
| 2 / 23 | Llama 3.1 70B Instruct · 70.6B | 27.9% | ~47 GB |
| 3 / 24 | Llama 3.2 90B Vision Instruct · 88.6B | 27.3% | ~58 GB |
| 4 / 25 | Llama 3.3 70B Instruct · 70.6B | 23.0% | ~47 GB |
| 5 / 27 | DeepSeek R1 Distill Qwen 32B · 32.8B | 19.5% | ~21 GB |
| 6 / 29 | Mistral Nemo Instruct 2407 · 12.2B | 17.6% | ~8.1 GB |
| 7 / 31 | Llama 3.2 11B Vision Instruct · 10.7B | 16.8% | ~7.0 GB |
| 8 / 32 | Qwen2.5 72B Instruct · 72.7B | 16.2% | ~45 GB |
| 9 / 33 | Llama 3.1 8B Instruct · 8.0B | 15.1% | ~5.3 GB |
| 10 / 35 | Qwen2 VL 72B Instruct · 73.4B | 12.8% | ~45 GB |
| 11 / 36 | Phi 4 · 14.7B | 11.6% | ~9.5 GB |
| 12 / 37 | Llama 3.2 3B Instruct · 3.2B | 10.1% | ~2.1 GB |
| 13 / 38 | Qwen2.5 7B Instruct · 7.6B | 7.8% | ~5.0 GB |
| 14 / 39 | Llama 3.2 1B Instruct · 1.2B | 6.6% | ~0.8 GB |
| 15 / 40 | Qwen2 VL 7B Instruct · 8.3B | 3.7% | ~5.4 GB |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Llama 3.2 1B Instruct, 1B, score 6.6% — on the efficiency frontier (best score at its size or smaller).
- Llama 3.2 3B Instruct, 3B, score 10.1% — on the efficiency frontier (best score at its size or smaller).
- Llama 3.1 8B Instruct, 8B, score 15.1% — on the efficiency frontier (best score at its size or smaller).
- Llama 3.2 11B Vision Instruct, 11B, score 16.8% — on the efficiency frontier (best score at its size or smaller).
- Mistral Nemo Instruct 2407, 12B, score 17.6% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek R1 Distill Qwen 32B, 33B, score 19.5% — on the efficiency frontier (best score at its size or smaller).
- Llama 3.1 70B Instruct, 71B, score 27.9% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek R1, 684B, score 34.9% — on the efficiency frontier (best score at its size or smaller).
BALROG: frequently asked questions
- What is the best open LLM on BALROG?
- DeepSeek R1 is the top open model on BALROG, scoring 34.9%. Among all models tested — including proprietary ones — it ranks #14. The top model overall is GPT 6 Astra Max (OpenAI) at 68.3%.
- What's the best BALROG model you can run on a 24 GB GPU?
- Mistral Nemo Instruct 2407 is the highest-scoring open model that fits in 24 GB at 4-bit quantization (~8.1 GB), scoring 17.6% on BALROG.
- What's the best BALROG model you can run on a 12 GB GPU?
- Mistral Nemo Instruct 2407 is the highest-scoring open model that fits in 12 GB at 4-bit quantization (~8.1 GB), scoring 17.6% on BALROG.
- Can open models match proprietary models on BALROG?
- Not quite on BALROG: the strongest proprietary model (GPT 6 Astra Max) scores 68.3%, ahead of the best open model (DeepSeek R1) at 34.9% — 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.