Tongyi DeepResearch 30B A3B
Alibaba · 2025-09-16 · 30.5B parameters
Tongyi DeepResearch-30B-A3B is Tongyi Lab's open-weight, Qwen3-MoE-based agentic model for long-horizon information seeking. The released checkpoint has exactly 30,532,122,624 parameters; the technical report states 3.3B activated parameters per token only at rounded precision. It supports a 131,072-token context and both thinking and non-thinking generation via its chat template.
Benchmark scores
| Benchmark | Score |
|---|---|
| SnakeBench Average Apples | 2.9 |
| SnakeBench Best Apples | 13.0 |
| SnakeBench Rating | 25.7 |
| SnakeBench Total Apples | 58.0 |
| SnakeBench Win Rate | 52.6 |