Qwen3 Next 80B A3B Instruct
Qwen · released 2025-09-11 · apache-2.0 license
Qwen3-Next-80B-A3B-Instruct is Alibaba's efficiency-focused open-weight MoE (80B total, 3B active, Sept 2025, Apache-2.0) performing on par with the larger 235B-Instruct on several benchmarks: MMLU-Pro 80.6, GPQA 72.9, AIME25 69.5.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 81.32B total · MoE, — active |
| Architecture | qwen3_next |
| Context window | 262K tokens |
| Knowledge cutoff | 2025-09-30 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 72.9% |
|---|---|
| SWE-bench Verified | — |
| AIME | 69.5% |
| MMLU-Pro | 80.6% |
| BFCL v3 (tool use) | — |
| Composite score | 7.64 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 48.7 GB | 47.2 GB | — | 262K |
Strengths & weaknesses
Strengths: Highly efficient hybrid-attention ultra-sparse MoE (80B total, 3B active); Near-235B-Instruct quality at a fraction of active params; Strong long context (RULER 91.8)
Weaknesses: Falls short of the 235B on knowledge and some agentic tasks; Non-reasoning mode only