Skip to main content

​
Playground

An Experimental MoE LLM introducing DeepSeek Sparse Attention (DSA). It achieves dramatically lower long-context inference cost while maintaining V3.1’s high quality.

​
Key Features

  • Sparse Attention (DSA): Utilizes a novel sparse attention mechanism to cut inference costs significantly.
  • Cost Efficiency: Achieves near-parity with V3.1 quality at a fraction of the computational expense.
  • Long-Context Efficiency: Optimized for cost-effective handling of extremely large documents and code files.