Chinese AI developer DeepSeek has officially released version 3.1 of its artificial intelligence model, marking a notable advancement in its technology. The open-source model features a 685-billion-parameter architecture and expands the context window to 128,000 tokens, enabling it to process extensive information equivalent to a 300-page book in a single query. DeepSeek V3.1 integrates a hybrid inference system combining "Think" and "Non-Think" modes within one model, enhancing reasoning, coding, and tool-use capabilities. Benchmark results indicate improved performance across multiple domains, including coding (SWE-bench score rising from 45 to 66), multilingual reasoning (from 29 to 54), and search tasks (BrowseComp score increasing from 9 to 30), along with high scores on advanced math tests such as AIME 2024 (93.1) and GPQA (80.1). The model reportedly achieves a 71.6% accuracy on the Aider coding benchmark, surpassing GPT-4.5's 44.9% accuracy, while being more cost-effective. DeepSeek V3.1 unifies the previous V3 and R1 versions, offering faster response times and stronger agent skills through post-training improvements that boost multi-step tool use. The update is available across DeepSeek's official website, app, WeChat mini program, and API, with no changes to the API calling method. Despite a quiet release without a detailed model card, V3.1 quickly gained traction, ranking fourth on Hugging Face's trending list. DeepSeek has announced plans to raise API pricing for the V3.1 model and discontinue off-peak discounts starting September 6, 2025. Some analysts note that while V3.1 shows incremental intelligence gains and hybrid capabilities, it has not surpassed OpenAI's GPT-OSS 120B model. The release is viewed as a strategic step toward the agent era in AI, emphasizing integration of reasoning and quick task handling in a single architecture.