DeepSeek AI's V3.2-exp Slashes Long-Context AI Costs by 50%
DeepSeek AI, based in Hangzhou, has launched V3.2-exp, an experimental model designed to tackle the high inference costs in AI adoption. Unveiled on September 29, 2025, this model promises to cut API call prices by up to 50 percent in long-context cases, benefiting startups, universities, and nonprofits.
The central feature of V3.2-exp is Sparse Attention, a two-step process that reduces data processing volume. First, a 'lightning indexer' selects relevant input parts. Then, a 'fine-grained token selection system' identifies key tokens within those parts. This method nearly halves the costs for long-context operations. The model's efficiency results in a leaner, faster AI model that limits server strain while preserving accuracy, mirroring how people scan documents.
DeepSeek AI's latest release builds on its earlier work with the R1 model, which used reinforcement learning to reduce training expenses. This new model continues the company's focus on efficiency rather than chasing ever-larger models, offering an alternative strategy in the broader U.S.-China AI rivalry. The model is open-weight and available on Hugging Face, allowing independent researchers to test and verify its claims.
DeepSeek AI's V3.2-exp promises to make AI technology more accessible by significantly reducing inference costs, particularly for long-context tasks. This model, released on September 29, 2025, is open for independent verification and could pave the way for more cost-efficient AI solutions.
Read also:
- EPA Administrator Zeldin travels to Iowa, reveals fresh EPA DEF guidelines, attends State Fair, commemorates One Big Beautiful Bill
- Leaders at HIT Forum 2025: Middle Powers Key to Asia's Security
- JPMorgan Chase Announces Plans for a Digital Bank Launch in Germany's Retail Sector
- Derrick Xiong, one of the pioneers behind the drone company EHang