Google releases Gemini API map features; Tencent launches AI programmer Ada → Here’s the latest AI news!

AI Frontier Express ↓↓

01 Google opens up Gemini API’s map features, enabling developers to build next-generation location-aware AI applications.
Google announced the full-scale availability of map integration features for the Gemini API. Starting today, developers can use the API to access real-time geospatial data—covering 250 million locations worldwide—including addresses, business hours, ratings, and more. Leveraging Gemini’s natural language understanding capabilities, the API can automatically parse location-based queries and generate highly accurate responses. For example, AI can directly recommend nearby restaurants and display real-time traffic conditions, or it can plan multi-stop itineraries based on user preferences. This feature works seamlessly with Google Search: while Maps provides structured factual data, Search supplements it with dynamic information, significantly enhancing the completeness of responses. Developers can quickly integrate the feature via the Python SDK or Google AI Studio. The pricing is $25 per thousand factually grounded prompts. This capability is ideal for applications such as travel planning, local lifestyle services, and real estate recommendations, helping developers create immersive spatial intelligence experiences.

02 The domestically developed large language model MiniMax-M2 has been open-sourced and ranked among the global top five, with a cost only 1/12 of that of Claude 4.5, and its performance is on par with GPT-5.

On October 27, Shanghai-based AI company MiniMax open-sourced its next-generation large language model, MiniMax-M2. In the globally recognized benchmark Artificial Analysis, MiniMax-M2 ranked among the top five worldwide and emerged as the first Chinese open-source model to make it into this elite group. Its overall cost is only 8% of that of Claude 4.5 (about 1/12) and 1/7 of GPT-5, while its inference speed has nearly doubled. M2 excels in core tasks such as code generation and tool invocation, supporting automatic debugging across multiple files and seamless collaboration with complex toolchains. Within two weeks of its open-source release, global API calls have been free of charge. Industry experts commend MiniMax-M2 for achieving a breakthrough in “high intelligence plus low cost” through technological innovations like the CISPO loss function, and anticipate that it will help accelerate the widespread adoption of AI technologies.

03 Meituan has open-sourced LongCat-Video! It supports text-to-video, image-to-video, and video continuation, enabling 5-minute-long videos without any loss in quality—and paving the way for a new chapter in world-model exploration.

On October 27, Meituan’s LongCat team released and open-sourced the video generation model LongCat-Video. This model supports three core tasks—text-to-video (720p/30fps high-definition), image-to-video (strictly preserving the main subject and style of the image), and video continuation—using a unified architecture without requiring additional adaptation. Relying on pre-training for video continuation and the Block-Causal Attention mechanism, the model can stably generate long videos lasting up to 5 minutes, addressing industry pain points such as temporal discontinuities across frames and physically implausible motions. Its inference speed has been boosted by a factor of 10.1. In public benchmarks like VBench, the model achieved state-of-the-art (SOTA) overall performance, with a common-sense understanding score of 70.94%, ranking first. In the future, this technology will be integrated into scenarios such as autonomous driving and embodied intelligence, helping to build a bridge between the digital and physical worlds.

04 Tencent Launches AI Programmer Ada: End-to-End Automated Programming Pioneers a Revolution in Development Efficiency

On October 27, Tencent launched Ada, the industry’s first technology preview of an AI-powered programmer, which revolutionizes the software development process with four core capabilities. Users simply submit their requirements via WeChat Work, and Ada automatically handles the entire workflow—from code generation and environment deployment to go-live release—while also automatically fixing online defects, generating documentation, and producing unit-test code. At the 1024 Developer Festival in 2025, Ada successfully completed the development of a hidden Easter egg on the official website and a pixel-style game. Its intelligent code review and end-to-end automation capabilities have significantly boosted development efficiency. Currently, Ada is collecting feedback through a waiting list to further optimize its features, and it is poised to become a core tool for developers.

05 The AI Website-Building Revolution! Kunlun Wanwei’s Skywork Launches the Web Clone Feature, Enabling You to Build a Professional-Grade Website in Just 10 Minutes.

On October 27, 2025, Skywork AI, a subsidiary of Kunlun Wanwei, officially launched its Web Clone feature. Users simply need to provide a webpage link, upload a file, or enter a text description, and the AI will generate a fully functional, stylistically faithful web prototype within minutes. This feature uses advanced algorithms to analyze page logic and hierarchy at a structural level, enabling precise reconstruction of web pages. It also supports the generation of dynamic interactive elements and allows users to edit and adjust the output in real time. Enterprise users can instantly transform their company information into a fully-fledged official website, while content creators can generate dynamic webpages—including maps and images—simply by providing text descriptions, completely revolutionizing traditional website-building processes. This breakthrough significantly lowers the technical barrier and marks the official arrival of an AI-driven era where “an idea equals a website.”

06 Gemini Powers Google Earth! A Major Upgrade to Spatial Reasoning—U.S. Subscribers Soon to Unlock a New AI-Driven Geospatial Experience

Google Earth has announced the integration of the Gemini large language model, enabling cross-dimensional geographic analysis through natural language interactions. Users can now directly engage in conversations to query environmental risks—such as identifying infrastructure vulnerable to storm threats or tracking areas affected by algal blooms. The system automatically links weather forecasts, satellite imagery, and demographic data to generate comprehensive insights. Currently, this feature is available to users enrolled in the “Trusted Tester Program.” In the coming weeks, it will be rolled out to Google AI Pro/Ultra subscribers in the U.S. and to users of the Earth Professional Edition, accelerating the shift of geospatial information from specialized tools toward accessible, intelligent services for the general public.

07 AI Large Model “Reads Images to Identify Genes” in Just 1 Minute!

Lung cancer precision diagnosis and treatment achieves a groundbreaking breakthrough.

In October 2025, Kingmed Diagnostics, in collaboration with Tencent and the First Affiliated Hospital of Guangzhou Medical University, jointly announced the official launch of DeepGEM, a large-scale cancer pathology gene model. This model can predict driver gene mutations in lung cancer simply by analyzing routine pathology slide images, completing the test within one minute with an accuracy rate ranging from 78% to 99%. The cost is several times lower than that of traditional gene sequencing. Previous research has been published in The Lancet Oncology and has been verified using a sample size of 4,260 cases across 30 provinces nationwide. Currently, the model has been deployed on Kingmed Diagnostics’ platform, providing grassroots patients with an inclusive diagnostic and treatment pathway featuring “AI-based initial screening plus specific gene confirmation.” In the future, the three parties will expand the model to cover multiple cancers, including breast cancer, and develop a multimodal large-scale pathology gene model integrating multi-omics data, thereby ushering in a new era of “intelligent prediction” in cancer diagnosis and treatment.

08 Tesla’s “World Simulator” Revealed: AI Practices 500 Years of Driving in a Single Day; Robot-Brain Co-Evolution Shocks the Industry

At the 2025 top computer vision conference ICCV, Tesla officially unveiled the "World Simulator"—a virtual training system powered by neural networks. This system can generate highly realistic multi-camera driving scenarios based on real-world data, enabling Autonomous driving algorithms to accumulate driving experience equivalent to 500 years of human driving in just one day. At its core, the system adopts an end-to-end architecture that directly translates camera pixels into driving commands, thereby addressing the information loss inherent in traditional modular approaches. Moreover, it leverages a "Generative Gaussian Splash" technology to achieve interpretability of 3D scenes. Beyond being used for FSD iterations, this system is also simultaneously training Optimus robots to navigate complex environments such as factories, showcasing Musk's ambition to build a general-purpose AI engine. Currently, the massive amounts of data collected daily by Tesla’s fleet are being optimized through a simulator feedback loop, propelling Autonomous driving safety toward a level five times safer than human driving.

09 A Visual Revolution! DeepSeek Releases OCR Model Achieving 10x Text Compression, Processing 200,000 Document Pages in a Single Day

Recently, DeepSeek open-sourced a brand-new vision-language model called DeepSeek-OCR. Leveraging “contextual optical compression” technology, this model transforms text into images, achieving 10x lossless compression (with an OCR accuracy of 97%) and maintaining 60% accuracy even at 20x compression. At its core, the DeepEncoder architecture integrates SAM and CLIP, supporting multi-resolution inputs—from Tiny to Gundam mode. With just 100 visual tokens, it outperforms GOT-OCR2.0 (which requires 256 tokens), while with 800 tokens, it significantly surpasses MinerU2.0 (which demands over 6,000 tokens). A single A100 GPU can generate 200,000 pages of training data per day, supporting parsing of 100 languages as well as complex content such as charts, chemical formulas, and more. This breakthrough model offers a transformative solution for long-text processing and multimodal model training. The model has been open-sourced on GitHub, sparking intense discussion within the AI community about the “visual-is-compression” paradigm.

Source: School of Artificial Intelligence, AI Zhi Qian Yan

Editor: Yu Li

Proofreading: Zeng Xiong

Reviewed by: Zhang Weitian