“DeepSeek-V2 utilizes a sparse attention mechanism with a lightweight indexer that selectively attends to necessary tokens rather than all tokens, improving efficiency.”