DeepSeek-V3.2 Rivals Top AI Models at Fraction of the Cost With Breakthrough Sparse Attention Technology

DeepSeek-V3.2 is shaking up the AI industry, ranking fifth globally while costing a fraction of rival models, thanks to breakthrough sparse attention technology that slashes inference costs by converting attention complexity from quadratic to linear using only 37 billion of its 671 billion parameters per token.

• DeepSeek-V3.2 ranks fifth overall and second among open-weight models on the Artificial Analysis intelligence index, outperforming competitors like Grok 4 while costing just $0.28 per million input tokens and $0.48 per million output tokens.
• The model uses a sparse mixture-of-experts (MoE) architecture with 671 billion total parameters, activating only 37 billion per token, dramatically reducing computational costs while maintaining high performance across diverse benchmarks.
• A newly introduced DeepSeek Sparse Attention (DSA) mechanism builds on multi-head latent attention by using a 'lightning indexer' to selectively attend to only relevant preceding tokens, effectively converting attention complexity from quadratic to linear and slashing inference costs significantly.