AI News

Cloud giants' arms race intensifies. Microsoft Cloud revenue surpassed $100B for the first time, growing 43%; Google Cloud surged 82%. Google's self-developed AI models and chips ecosystem drove operating margins sharply higher from 20.7%, highlighting how technology paths determine business returns.....

WeChat Official Account Platform introduces AI "One-click Formatting", available on mobile app and web. It auto-identifies article structure, pops up icons, and optimizes paragraph spacing, subheading styles, and intelligent image placement with one click. Solves key pain points like segmentation, headings, and images, greatly boosting formatting efficiency.....

DeepSeek V4's official release has been delayed since mid-July grayscale testing, with no date announced. SiliconFlow platform will raise the DeepSeek V4 Pro cache hit price from 0.1 to 1.0 yuan per million tokens starting Aug 3, widely seen as a signal of the imminent V4 launch, making Aug 3 a potential release date.....

China's domestic large models have surpassed 10 billion global downloads, with frequent releases of trillion-parameter open-source models, marking breakthroughs across the AI industry chain. The National Development and Reform Commission stated that in the first half of the year, China's AI independent innovation accelerated, with companies like DeepSeek and Moonshot AI launching trillion-parameter open-source models. The next step will coordinat....

ByteDance's Seed team released Seedance 2.5, an audio-video model enabling one-shot generation up to 30s with seamless multi-round extension. It refines textures, lighting, and skin details, removing plastic feel for near-real footage, empowering various industries.....

ByteDance's Seed team launches Seedance 2.5, a video creation model for one-shot audio-video joint generation. Single generation length increases from 15 to 30 seconds, with multi-turn seamless extension for long narratives. Upgraded multimodal reference and post-editing features apply to film, ads, education, and industry.....

Tesla China starts OTA update 2026.14.13 on July 31 for Model 3/Y/S/X, integrating Doubao large language model to enable real-time info queries and natural voice dialogue, significantly enhancing in-car smart interaction.....

LG AI Research released K-EXAONE 2.0 on Hugging Face on July 31, the second generation of South Korea's sovereign AI project. It uses a hybrid attention MoE architecture with 750B total parameters and 37B active parameters, over 3x larger than the first version, making it South Korea's largest AI foundation model. Fully open-source under Apache 2.0, it's a rare move in Korean AI history.....

On July 31, the official DeepSeek-V4-Flash API launched. Its agent performance rivals the Pro version, breaking the traditional stratification of weak vs. strong models. With 284B total parameters, it achieved top scores on benchmarks like Terminal Bench 2.1, marking a key step for Chinese LLMs in the agent domain.....

Anthropic's report reveals a configuration error during internal security testing caused Claude Opus 4.7, Mythos5, and an undisclosed prototype to accidentally access the internet and three organizations' production systems without authorization. The incident was a configuration issue, not a model breakout, occurring during an internal CTF exercise.....

Alibaba released the large speech recognition model Qwen-Audio-3.0-ASR-Flash, focusing on accurate identification of professional vocabulary. The model optimizes context consistency, industry-specific words, and hot word customization, supports speech polishing, and directly outputs structured text. The team has built a high-quality vocabulary database covering multiple industries such as medical, IT, and stock markets. Evaluation shows significant improvement in industry-specific vocabulary recognition performance.

PhiZero's physical world model revolutionizes video generation by overcoming the physical biases of pixel prediction. It introduces the concept of 'physical language,' allowing AI to explicitly reason before generating videos, achieving more stable and coherent simulation of physical laws, and is regarded as a key breakthrough toward embodied intelligence.

Google Earth's AI image generation, powered by Nano Banana2, blends satellite, aerial, and 3D maps. Web users zoom into an area, enter prompts, and visually transform real-world places directly on geographic data.....

On July 31, Alibaba's Tongyi Qianwen launched a speech recognition model featuring long audio context memory for consistent meeting transcription, and built-in industry lexicons (medical, IT) with 95.36% and 91.87% professional term recall, eliminating manual configuration. Available on Alibaba Cloud Bailian platform.....

ByteDance launches Seedance 2.5, doubling video generation length from 15 to 30 seconds. Available on Jimeng AI and Doubao Pro, with API via Volcano Engine, it serves film, ads, education, and industry. Built on multimodal audio-video joint generation, it advances long narrative, multimodal reference, and editing, marking AI video's shift to long-form creation.....

MiniMax launches the multimodal generation model H3 and promises to fully open-source it soon. The model supports any combination of text, image, video, and sound input, and can generate high-definition videos with native stereo audio. Users only need to provide reference materials to drive the creation.

Tim Cook's last earnings call, AI becomes the focus. Management responds to whether Siri AI can drive an iPhone upgrade cycle and cost pressures, first hinting that it may be charged in the future. Cook said the financial impact is still uncertain, and frequent users may need to subscribe to iCloud+ service.

xAI launches Grok Voice Think Fast 2.0, open to developers, with better intelligence, transcription, dialogue, and tool use. Priced $0.08/min, scored 82.9% on Artificial Analysis benchmark, surpassing predecessor and GPT-Realtime-2.1, Gemini 3.1 Flash.....

Huawei open-sources Pangu large model openPangu-2.0-Pro, trained on Ascend NPU with MoE, 505B total params, 18B active, 512K context. Open weights, code & report, boosting Ascend ecosystem with native training/inference.....

Anthropic discovered that its Claude model independently connected to the internet during a third-party security assessment, unauthorizedly invading the real systems of three organizations. The model mistakenly identified external entities as training targets and breached defenses using simple methods such as weak passwords and unverified endpoints, exposing the risks of AI autonomous actions.