CLEM Clemscore: leaderboard
CLEM text release v3.0: models play 17 text conversational games spanning instruction following, language grounding, dialogue, negotiation, navigation and planning. 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: Clemscore (%), text v3.0. Source: huggingface.co. Saturation forecast: Around January 2027. 31 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (High) | 84.19 | #110 (GPT-5.2) |
| 2 | Gemini 3 Flash (Preview) | 84.03 | #82 |
| 3 | GPT-5.2 (Medium) | 79.61 | #110 (GPT-5.2) |
| 4 | GLM-4.7 | 78.05 | #192 |
| 5 | Kimi K2 (Thinking) | 77.79 | #243 (Kimi K2) |
| 6 | GLM-4.6 | 63.91 | #252 |
| 7 | Kimi K2.5 (Non-reasoning) | 60.28 | #144 (Kimi K2.5) |
| 8 | Qwen 3 Max | 59.66 | #208 |
| 9 | DeepSeek V3.2 | 59.61 | #205 |
| 10 | GLM-5 (Non-reasoning) | 58.68 | #141 (GLM-5) |
| 11 | MiniMax-M2.5 | 55.68 | #303 |
| 12 | DeepSeek V3.2 (Non-reasoning) | 52.94 | #205 (DeepSeek V3.2) |
| 13 | Llama 3.3 70B Instruct | 50 | #534 |
| 14 | Qwen 2.5 72B Instruct | 48.07 | #454 |
| 15 | Llama 3.1 70B Instruct | 46.8 | #561 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=clem-clemscore · How It Works · Data refreshed daily, snapshot 2026-10-20.