Chinese AI startup DeepSeek has officially released version 3.1 of its artificial intelligence model, marking a step toward what the company calls the "agent era." The DeepSeek V3.1 model features a hybrid inference architecture that integrates two modes—"Think" and "Non-Think"—within a single model. This design enables faster reasoning and improved efficiency compared to its predecessor, DeepSeek-R1-0528. The model supports a significantly expanded context window of 128,000 tokens, allowing it to process large volumes of information in a single query, equivalent to a 300-page book. With 685 billion parameters, DeepSeek V3.1 combines chat, reasoning, and coding capabilities into one architecture and demonstrates improved performance on several benchmarks, including a rise in the SWE-Bench coding score from 44.6% to 66% and enhanced multilingual reasoning and search capabilities. Post-training optimizations have strengthened the model's tool use and multi-step agent task handling. The update maintains compatibility with the existing API and is available on multiple platforms, including the official website, app, WeChat mini program, Hugging Face, and Lightning's Model API. DeepSeek V3.1 is optimized for domestic Chinese AI chips, signaling progress toward semiconductor self-sufficiency amid ongoing geopolitical tensions. This has contributed to a surge in Chinese semiconductor stocks, notably Cambricon Technologies, whose shares have more than doubled in 2024, reaching a market capitalization exceeding RMB 500 billion ($70 billion). Analysts view DeepSeek's advancements and chip optimizations as part of China's broader strategy to enhance its AI and semiconductor industries. While the update is considered incremental rather than revolutionary, DeepSeek V3.1 is regarded as competitive with leading global AI models in reasoning efficiency and agent capabilities.