Tool-calling on ACEBench English
25Multistep (Agent)Qwen3-4B-RL (ours)
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-4B-RL (ours)Source=Open-source2026.05 | 25 | 30 | 90 | 79.5 | 80 | 70 | 80 | 72 | 90 | 92 | 77.5 | |
| Nova-1-ProSource=Closed-source2026.05 | 20 | 56.7 | 94.3 | 81.5 | 88.5 | 74 | 76 | 62 | 44 | 84 | 81.4 | |
| Nova-1-LiteSource=Closed-source2026.05 | 20 | 60 | 86 | 60 | 78.5 | 68 | 74 | 14 | 46 | 78 | 73.4 | |
| Gemma3-27B-InstructSource=Open-source2026.05 | 20 | 56.7 | 88.3 | 55 | 81 | 66 | 76 | 70 | 32 | 90 | 76.8 | |
| Magistral-Small-24B-2509Source=Open-source2026.05 | 20 | 50 | 91 | 58 | 82 | 70 | 74 | 40 | 18 | 48 | 67.4 | |
| Claude-4-SonnetSource=Closed-source2026.05 | 15 | 53.3 | 94 | 79 | 86.5 | 76 | 86 | 96 | 76 | 94 | 81.8 | |
| Qwen3-4B (base)Source=Open-source2026.05 | 10 | 20 | 77.3 | 64.7 | 66.5 | 64 | 76 | 58 | 84 | 74 | 65.4 |