AI With Me Blog & News

The release of DeepSeek-V3.2 marks a shift in the focus of competition among large-scale models towards "efficiency." Key highlights include: 1) Innovative architecture: Utilizing DSA sparse attention, significantly improving efficiency for long text processing; 2) Dramatic cost reduction: API call costs have decreased by up to 75%; 3) Dual-model strategy: Providing a balanced V3.2 version and a Speciale version dedicated to complex inference; 4) Open source: The model has been open-sourced on platforms such as Hugging Face, promoting technology accessibility.

On November 27, 2025, DeepSeekAI officially released the DeepSeekMath-V2 model, a large language model focused on mathematical reasoning. This model aims to address a key challenge in current AI mathematical reasoning: ensuring the correctness and rigor of the reasoning process, rather than simply pursuing the accuracy of the final answer. This article will introduce the main functions, technical principles, performance, target audience, and commercial potential of DeepSeekMath-V2 based on information from the official GitHub repository and the HuggingFace model library.

Anthropic officially released Claude Opus 4.5 today (API model name: `claude-opus-4-5-20251101`), the latest iteration of its flagship model series.

In November 2025, Meta announced the research progress of its 3D world generation system, WorldGen, on its official blog. This system can generate freely explorable 3D scenes from text descriptions (such as "cartoon-style medieval village") and is directly compatible with mainstream engines such as Unity and Unreal Engine. This article will provide an objective analysis from the perspectives of implementation principles and applicable scenarios, based on official technical documentation.

Google's Nano Banana Pro, released on November 20, 2025, is far more than a simple model iteration; it's a landmark event marking the industrialization of AI image generation. Deeply integrating the powerful inference capabilities of the Gemini 3 Pro, it achieves a paradigm shift from "random creative generation" to "precise and controllable production." This update represents a significant breakthrough in 4K ultra-high resolution, complex instruction adherence, multilingual text rendering, and cross-image character consistency. Its significance lies in upgrading AI image generation from a creative toy into a reliable productivity engine, directly empowering professional fields such as design, advertising, and film.

On November 20, 2025, OpenAI quietly but strategically released two models, GPT-5.1 Pro and GPT-5.1-Codex-Max. This move is widely seen as a direct response to Google's earlier release of Gemini 3 Pro, marking a new phase in the large-scale model race.

On November 18th, Google officially released its new flagship AI model, Gemini 3.0 Pro. This is not a routine upgrade, but an evolution from a passive tool to an active agent. On the LMArena global leaderboard, Gemini 3.0 Pro topped the list with an Elo score of 1501, surpassing all its previous competitors.

At a crucial juncture in the transition of AI technology from the laboratory to large-scale application, Alibaba officially launched the "1000 Questions" project for consumers on November 17, 2025. This initiative marks a significant leap in Alibaba's AI strategy from the B2B to the B2C market.

Say goodbye to information overload and the illusion of AI. Let Google NotebookLM become your personal research assistant and usher in a new era of efficient scientific research.