CLEM Hot Air Balloon: leaderboard
CLEM text release v3.0: models coordinate decisions in a constrained balloon scenario. Scores combine protocol-valid play and conditional task quality, higher is better. Headline preserves the native published product of macro-average played percentage and macro-average conditional quality; per-game scores multiply source-published played and quality components without extra rounding. Source temperature, player configuration and exact evidenced model/checkpoint settings are retained. Undefined conditional quality remains missing rather than imputed zero.
Metric: Game Clemscore (%), text v3.0; published played × quality / 100. Source: huggingface.co. Saturation forecast: Around January 2027. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-4.7 | 90.89 | #192 |
| 2 | GLM-4.6 | 90.16 | #252 |
| 3 | Kimi K2 (Thinking) | 89.26 | #243 (Kimi K2) |
| 4 | DeepSeek V3.2 | 86.62 | #205 |
| 5 | GLM-5 (Non-reasoning) | 75.02 | #141 (GLM-5) |
| 6 | GPT-OSS-20B | 73.82 | #518 |
| 7 | MiniMax-M2.5 | 72.8 | #303 |
| 8 | Gemini 3 Flash (Preview) | 70.69 | #82 |
| 9 | Kimi K2.5 (Non-reasoning) | 70.29 | #144 (Kimi K2.5) |
| 10 | GPT-OSS-120B | 59.14 | #341 |
| 11 | GPT-5.2 (High) | 58.06 | #110 (GPT-5.2) |
| 12 | Qwen 2.5 72B Instruct | 48.71 | #454 |
| 13 | Llama 3.3 70B Instruct | 47.74 | #534 |
| 14 | GPT-5.2 (Medium) | 36.9 | #110 (GPT-5.2) |
| 15 | Qwen 2.5 Coder 32B Instruct | 33.36 | #528 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=clem-hot-air-balloon · How It Works · Data refreshed daily, snapshot 2026-10-20.