DeepSeek V4-Pro
DeepSeek V4-Pro is a 1.6T-parameter Mixture-of-Experts model with 49B active parameters per token, hybrid attention and 73% lower per-token inference FLOPs vs V3.2.
Key features
DeepSeek V4-Pro · AI Models
Parameters
1.6T total · 49B active
Context window
1M tokens
Released
April 24, 2026
Pricing
DeepSeek API
Key features
- 1.6T total / 49B active parameters MoE.
- Hybrid attention combining Compressed Sparse and Heavily Compressed Attention.
- 73% reduction in per-token inference FLOPs vs DeepSeek V3.2.
- 90% reduction in KV-cache memory footprint.
Best for
Pick DeepSeek V4-Pro for the highest DeepSeek quality on hard reasoning, large-context workloads and self-hosted enterprise inference.
Frequently Asked Questions
What replaced DeepSeek V3?
DeepSeek V4 replaces V3. V3 and V3.2 are being retired after July 24, 2026.
Open Chat
DeepSeek V4-Pro is a 1.6T-parameter Mixture-of-Experts model with 49B active parameters per token, hybrid attention and 73% lower per-token inference FLOPs vs V3.2.