DeepSeek Launches V3.2-Exp With Sparse Attention, Paving Path to Next-Gen AI

San Francisco / September 29, 2025 — DeepSeek has unveiled V3.2-Exp, an experimental model that introduces DeepSeek Sparse Attention (DSA). Built on top of V3.1-Terminus, the new release is framed as a bridge toward a next-generation architecture.

The company says DSA improves efficiency for both training and inference while extending the ability to process longer sequences. Analysts note that handling long context windows has become a central competitive challenge in 2025, and DeepSeek’s approach could lower costs for developers working with complex, memory-heavy workloads.

V3.2-Exp is available immediately through the DeepSeek app and API. Pricing changes have been hinted at but not fully detailed, with developers encouraged to check the API console for current rates. The firm has called this an interim release, signaling larger architectural updates to come.

Suggested Stories

• DeepSeek-V3.1 Launches Hybrid Reasoning in One Model — Context on how the previous release set the stage for V3.2-Exp.

• OpenAI’s Realtime API Reaches General Availability with gpt-realtime — Shows how rivals are pushing real-time AI capabilities in parallel.

• Tesla Launches “Hey Tesla” Voice AI in China with DeepSeek and ByteDance — Links DeepSeek’s growing partnerships and applied use cases.

• NVIDIA-Backed UK AI Firm nScale Secures $1.1B in Funding to Expand Global Reach — Highlights competitive momentum in global AI infrastructure.

• UN Moves to Launch Global AI Governance Dialogue amid Rising Risks — Places DeepSeek’s experimental releases in the wider policy and governance debate.