DeepSeek V4.1 Flash: New Architecture, Native Multimodal, and the End of V4 Pro
DeepSeek officially released V4.1 Flash on September 10, 2026. It's the smallest model in their new architecture family—and it outperforms the V4 Pro it replaces.
New Architecture: Causal Encoder-Decoder
V4.1 Flash uses an asymmetric Causal Encoder-Decoder design:
| Component | Active Parameters |
|---|---|
| Input | 8B |
| Output | 16B |
Total parameters: 552B MoE.
This asymmetry reduces inference cost while maintaining capability.
KV Cache Optimization
The practical impact for Agent workloads:
| Metric | Reduction vs. Previous Gen |
|---|---|
| HBM demand | 1/4 |
| SSD storage | 1/8 |
Cache-hit costs dominate agent task economics. This directly addresses that.
Multimodal Consolidation
The standalone Vision model is retired. Native multimodal support is now integrated into V4.1 Flash:
# Old: separate vision endpoint
model = "deepseek-v4-flash-vision-exp" # retired
# New: unified model
model = "deepseek-flash"
Available across API, Web Chat, and App.
V4 Pro Retirement
V4.1 Flash beats V4 Pro on performance, cost, speed, and total runtime. DeepSeek's official statement: benchmark results are "ahead of flagship models, including DeepSeek-V4-Pro."
Retirement Timeline
| Date | Action |
|---|---|
| Now | V4 Pro retired from Web Chat & App |
| Sept 14, 2026 (04:00 UTC) | All deepseek-v4-pro API requests route to V4.1 Flash at Flash rates |
| Until V4.1 Pro launches | Routing continues |
# Before: explicit Pro model
model = "deepseek-v4-pro" # retired
# After: routes to V4.1 Flash automatically
model = "deepseek-flash"
Infrastructure Context
DeepSeek is simultaneously hiring ~150 senior backend engineers to overhaul infrastructure strained by growing user demand and agent workloads. The new architecture reduces backend pressure; the hiring addresses the rest.
Key Takeaways
- V4.1 Flash replaces both V4 Pro and the standalone Vision model
- New architecture cuts cache costs dramatically
- Native multimodal is now standard, not separate
- API model name:
deepseek-flash
What are your thoughts on the architecture shift? Testing V4.1 Flash in your workflows?
DeepSeek official Account:
DeepSeek official changelog:
DeepSeek official Documentation:



Top comments (0)