DeepSeek V3.2
deepseek-ai · released 2025-12-01 · mit license
DeepSeek-V3.2 is an open-weights MoE (685B total, 37B active, Dec 2025) pairing sparse attention with strong reasoning: 82.4% GPQA Diamond, 70% SWE-bench Verified (vendor). Artificial Analysis rates the non-reasoning variant 32.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 685.4B total · MoE, — active |
| Architecture | deepseek_v32 |
| Context window | 164K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 82.4% |
|---|---|
| SWE-bench Verified | 70% |
| AIME | 94.17% |
| MMLU-Pro | — |
| BFCL v3 (tool use) | — |
| Composite score | 7.23 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 399 GB | 397.5 GB | — | 164K |
Strengths & weaknesses
Strengths: Strong competition math (AIME 94.2); Efficient long-context via DeepSeek Sparse Attention; Open weights (MIT)
Weaknesses: AA Intelligence Index (32) trails top proprietary; Weak on long-horizon agentic tasks (APEX 7)