DeepSeek V4 Flash (Preview)
Weights availableDeepSeek V4 Flash (Preview) weights published by DeepSeek in its Hugging Face organization (deepseek-ai/DeepSeek-V4-Flash). Self-hosting costs and serving limits depend on your deployment.
Access details: DeepSeek open weights (Hugging Face)
The Hugging Face repository. Download the weights and run them yourself or with a host.
deepseek-ai/DeepSeek-V4-FlashUpdated Oct 1, 2026
Capabilities & limits
- Context window
- 1M
- Maximum output
- Not published
- Image input
- Unsupported
- Tool calling
- Supported
- Structured output
- Not published
Hosted price
USD / 1M tokens- Input
- $0.04186
- Output
- $0.08372
What OpenRouter charges for deepseek/deepseek-v4-flash. Other hosts set their own prices.
Free to download. Running the weights costs whatever your hardware or host charges.
Price history
Oct 1, 2026
Input Not applicable per million tokens
Oct 1, 2026
Output Not applicable per million tokens
Technical details
- Parameters
- 284B
- Active parameters
- 13B
- Architecture
- Mixture-of-Experts with hybrid Compressed Sparse Attention and Heavily Compressed Attention, manifold-constrained hyper-connections and the Muon optimizer
- Input modalities
- Text
- Output modalities
- Text
License
MIT
From DeepSeek
Official announcement ↗How they describe it
DeepSeek-V4-Flash with 284B parameters (13B activated)
What they say is new
Reasoning capabilities closely approach V4-Pro.
Release & lifecycle
- Published
- Apr 24, 2026
- available
- As of Oct 1, 2026