DeepSeek · 284B (13B active) · Mixture of Experts
Agentic MoE with speculative decoding — outperforms V4-Pro on coding benchmarks despite far fewer active parameters
868.6K downloads
2.9K likes
2026-07 1024K context
Use Cases
chat code reasoning
Mixture of Experts
Total experts: 256
Active experts: 6
Active params: 13.0B
| Quant | Bits | VRAM | Quality | Status |
|---|---|---|---|---|
| Q2_K | 2 | 91.4 GB | low | — |
| Q3_K_M | 3 | 127.8 GB | moderate | — |
| Q4_K_M | 4 | 146 GB | good | — |
| Q5_K_M | 5 | 182.3 GB | good | — |
| Q6_K | 6 | 218.7 GB | excellent | — |
| Q8_0 | 8 | 291.4 GB | excellent | — |
| F16 | 16 | 582.4 GB | lossless | — |
GPUs that run DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 needs about 146 GB of VRAM at Q4_K_M, so it runs comfortably on a GPU with 192 GB of VRAM or more (3 of the GPUs we track). Common cards, smallest first:
About this model
Agentic MoE with speculative decoding — outperforms V4-Pro on coding benchmarks despite far fewer active parameters