DeepSeek V4 Pro
deepseek-ai · released 2026-04-24 · mit license
DeepSeek-V4-Pro is a 1.6T-param (49B active) open-weight MoE with 1M context (April 2026), rated the top open model on Artificial Analysis's index (52). NIST/CAISI's independent eval found it trails the frontier and its own vendor numbers.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 861.61B total · MoE, — active |
| Architecture | deepseek_v4 |
| Context window | 1049K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 90.1% |
|---|---|
| SWE-bench Verified | 80.6% |
| AIME | — |
| MMLU-Pro | 87.5% |
| BFCL v3 (tool use) | — |
| Composite score | 7.33 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 501.2 GB | 499.7 GB | — | 1049K |
Strengths & weaknesses
Strengths: Top open-weights model on the Artificial Analysis Intelligence Index at release (52); Top-tier coding: LiveCodeBench 93.5, SWE-bench Verified 80.6 (vendor); 1M-token context, MIT licensed
Weaknesses: NIST/CAISI places real-world capability behind the US frontier; Self-reported scores run higher than independent held-out evals