DeepSeek, a China-based AI research firm, has released an experimental language model called V3.2-exp aimed at reducing inference costs for long-context tasks.

The model introduces a system known as DeepSeek Sparse Attention, which uses a “lightning indexer” to select key excerpts and a separate module for fine-grained token selection, allowing the model to process lengthy inputs with lower server loads.

According to DeepSeek, initial tests suggest API call costs could be halved for long-context operations, though further independent evaluation is needed.

V3.2-exp is open-weight and available on Hugging Face, enabling third parties to conduct their own assessments.

DeepSeek previously launched its R1 model earlier this year and continues to explore ways to make transformer-based AI models more efficient.

🔗 Source: TechCrunch

Recent DeepSeek developments