DeepSeek V4 Flash
deepseek-ai · released 2026-04-24 · mit license
DeepSeek-V4-Flash is the lighter 284B (13B active) member of the V4 family with 1M context (April 2026, MIT). In max-reasoning mode it approaches the Pro on coding, lagging on knowledge and the hardest agentic work.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 158.07B total · MoE, — active |
| Architecture | deepseek_v4 |
| Context window | 1049K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 88.1% |
|---|---|
| SWE-bench Verified | 79% |
| AIME | — |
| MMLU-Pro | 86.4% |
| BFCL v3 (tool use) | — |
| Composite score | 7.09 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 93.2 GB | 91.7 GB | — | 1049K |
Strengths & weaknesses
Strengths: 284B total / 13B active MoE with 1M context, very efficient; Strong coding for its active size (LiveCodeBench 91.6); MIT licensed
Weaknesses: Knowledge and hardest agentic tasks trail the Pro variant; Scores depend heavily on reasoning-effort mode; vendor-reported only