QwenClawBench final leaderboard
The published rows below are from the QwenClawBench leaderboard. The GLM-5.2 and DeepSeek V4 Flash rows are the final strict results from this evaluation, with invalid samples counted as zero.
| Reference position | Model | Score | Auto | LLM |
|---|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 60.5 | 74.5 | 73.4 |
| 2 | Claude Opus 4.6 | 59.5 | 75.8 | 72.1 |
| reference projection 3 | GLM-5.2 (this evaluation) | 59.26 | 72.50 | 69.10 |
| 3 | GLM-5.1 | 58.7 | 74.6 | 70.9 |
| 4 | Qwen3.6-Plus | 57.4 | 74.7 | 69.0 |
| 5 | GPT-5.4high | 56.7 | 71.5 | 72.9 |
| 6 | MiMo-V2-Pro | 56.5 | 73.0 | 69.1 |
| 7 | Claude Opus 4.5high | 56.1 | 75.1 | 70.7 |
| 8 | GLM-5 | 54.9 | 72.9 | 66.7 |
| 9 | Qwen3.5-Plus | 53.0 | 71.0 | 65.8 |
| 10 | Kimi-K2.5 | 51.9 | 69.5 | 62.6 |
| 11 | DeepSeek-V3.2-Thinking | 50.7 | 71.2 | 62.4 |
| 12 | MiniMax-M2.7high | 50.5 | 70.6 | 65.1 |
| below published rows | DeepSeek V4 Flash (this evaluation) | 46.52 | 60.99 | 55.84 |
Evaluated results
Each evaluated model covers 100 tasks with three OpenClaw rollouts per task. Invalid samples count as zero.
| Model | Samples | Valid samples | Completion rate | Score | Auto | LLM |
|---|---|---|---|---|---|---|
glm-5.2 | 300 | 282 | 94.00% | 59.2645 | 72.4988 | 69.0950 |
bailian/deepseek-v4-flash | 300 | 245 | 81.67% | 46.5202 | 60.9912 | 55.8392 |
Published-score comparison
| Evaluated model | Published reference | Score delta | Auto delta | LLM delta |
|---|---|---|---|---|
glm-5.2 | GLM-5.1 | +0.56 | -2.10 | -1.81 |
glm-5.2 | GLM-5 | +4.36 | -0.40 | +2.40 |
bailian/deepseek-v4-flash | DeepSeek-V3.2-Thinking | -4.18 | -10.21 | -6.56 |
The two evaluated models are reference projections rather than official rankings. The official leaderboard uses Claude as judge; this evaluation uses GLM-5.2. The model, harness, endpoint, and service versions also differ.