Back to News
DeepSeek AI•June 27, 2026
DeepSeek V4 Pro DSpark
Open Source
LLM
MoE
Long-Context
Open-Source
### TL;DR
DeepSeek V4 Pro DSpark is an advanced mixture-of-experts (MoE) language model optimized for high-efficiency, million-token context intelligence. It features a specialized speculative decoding module to accelerate inference performance while maintaining state-of-the-art reasoning and agentic capabilities.
Key Insights & Metrics
Pricing
Free
Cost structure
Version
1.0
Current release version
Hardware
A multi-node enterprise cluster equipped with at least 1 to 3.5 terabytes of VRAM, typically spanning 8 to 16 high-end datacenter GPUs like NVIDIA H100s or H200s
Compute requirements
Category
Open Source
Licensing model
Region
China
Primary region
Key Features
- 1.6 Trillion parameter MoE architecture (49B activated)
- Support for 1 million token context length
- Hybrid Attention Architecture (CSA and HCA) for efficiency
- Speculative decoding module integration
- Supports multiple reasoning effort modes (Non-think, High, Max)
- Optimized using the Muon optimizer for faster convergence
Discussion
0
Upvotes
0
Downvotes
0 reviews
Sign in to leave a review
Reviews
No reviews yet. Be the first to review!