According to @DeepLearningAI, DeepSeek's new 685B Mixture-of-Experts (MoE) AI model introduces a token-attention mechanism that processes only the most relevant tokens, enabling 2–3× faster long-context inference and reducing processing costs by 6–7× compared to its previous V3.1 model (source: DeepLearning.AI Twitter, Oct 22, 2025). The v3.2 model features MIT-licensed weights and API pricing of $0.28/$0.028/$0.42 per 1M input/cached/output tokens, promoting open-source adoption. It is specifically optimized for Huawei and other domestic Chinese chips, addressing hardware compatibility for the local market. While performance closely matches V3.1 overall, there are modest gains in coding and agentic tasks and minor trade-offs in science and math workloads, presenting new business opportunities for AI providers targeting cost-sensitive or China-centric deployments (source: DeepLearning.AI, The Batch).