Qwen3 30B A3B Instruct 2507
Qwen · released 2025-07-29 · apache-2.0 license
Qwen3-30B-A3B-Instruct-2507 is Alibaba's updated non-thinking 30.5B (3.3B active) open-weight MoE (July 2025, Apache-2.0): GPQA 70.4, AIME25 61.3, MMLU-Pro 78.4, strong long context. Artificial Analysis index 15.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 30.53B total · MoE, — active |
| Architecture | qwen3_moe |
| Context window | 262K tokens |
| Knowledge cutoff | 2025-06-30 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 70.4% |
|---|---|
| SWE-bench Verified | — |
| AIME | 61.3% |
| MMLU-Pro | 78.4% |
| BFCL v3 (tool use) | — |
| Composite score | 7.25 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 19.2 GB | 17.7 GB | — | 262K |
Strengths & weaknesses
Strengths: Strong general instruction-following and alignment; Good math for a non-thinking small MoE (AIME25 61.3); Excellent long context (RULER 86.8), 262K native
Weaknesses: Non-thinking; weaker on hard agentic tool-use; Aider-Polyglot coding (35.6) lags larger models