Qwen3 Coder 480B A35B Instruct
Qwen · released 2025-07-23 · apache-2.0 license
Qwen3-Coder-480B-A35B-Instruct is Alibaba's flagship open-weight agentic coding MoE (480B total, 35B active, July 2025, Apache-2.0). Independent leaderboards record SWE-bench Pro 38.7% (Scale AI) and Terminal-Bench 2.0 23.9% (tbench.ai).
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 480.15B total · MoE, — active |
| Architecture | qwen3_moe |
| Context window | 1049K tokens |
| Knowledge cutoff | 2025-06-30 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | — |
|---|---|
| SWE-bench Verified | — |
| AIME | — |
| MMLU-Pro | — |
| BFCL v3 (tool use) | — |
| Composite score | — |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 280 GB | 278.5 GB | — | 1049K |
Strengths & weaknesses
Strengths: Strongest open-weight agentic coding model at release (Qwen claims SWE-bench Verified SOTA among open models); Native 256K context extendable to 1M; Purpose-built agentic tool-calling
Weaknesses: Modest absolute scores on hardest independent suites (SWE-bench Pro 38.7); Very large (480B), costly to self-host