Qwen3-30B-A3B-Base
Alibaba · 2025-04-29 · 30.5B parameters
Qwen3-30B-A3B-Base is the pre-trained (not instruction-tuned) Mixture-of-Experts checkpoint of Alibaba's Qwen3 family, released April 29, 2025 under Apache 2.0 alongside the post-trained Qwen3-30B-A3B. It has 30,532,122,624 parameters with 128 experts, 8 routed per token (about 3.3B activated). Qwen describes a three-stage pre-training on about 36 trillion tokens in 119 languages, ending with long-context training at 32,768 tokens.
Benchmark scores
| Benchmark | Score |
|---|---|
| BBH (BIG-Bench Hard) | 81.5 |
| Capability | 120.7 |
| GSM8K | 91.8 |
| MATH | 59.0 |
| MGSM | 79.1 |
| MMLU | 81.4 |
| MMLU Pro | 61.5 |
| MMLU-Redux | 81.2 |
| SuperGPQA | 35.7 |