AI News

DeepSeek launches DeepSeek-V4 preview series as open-source, achieving million-character ultra-long context and leading domestic Agent collaboration and knowledge reasoning. Offers Pro (1.6T parameters, 49B activated) and Flash versions; Pro matches top closed-source models, Flash balances efficiency.....

DeepSeek launched its new flagship model V4 with a segmented strategy, introducing Flash and Pro versions for lightweight high-frequency applications and complex reasoning tasks, respectively. This move meets diverse scenario needs and reshapes AI commercialization benchmarks with competitive pricing, integrating the original deepseek-chat and deepseek-reasoner models.....

Wondershare's diagramming software, 'Wondershare MindMap,' has launched AI features, becoming China's first professional mind mapping tool to support AI Agent. It enhances knowledge management and office efficiency while retaining hand-drawing mode, currently in public beta.....

Soul AI Lab open-sourced SoulXFlashTalk, the first 1.4 billion-parameter real-time digital human generation model, featuring sub-second latency and 32 FPS, offering a complete real-time interaction solution. The open-source release includes project page, technical report, source code, and model weights, lowering industry R&D barriers.....

Tesla integrates ByteDance's Doubao large model for in-car voice services and DeepSeekChat for AI interaction, adopting a dual-model approach to localize in-car AI technology. This update follows the August 2025 partnership with Volcano Engine, strengthening Tesla's position in China's active market.....

Anthropic's Claude Mythos AI model, too powerful for public release due to cybersecurity capabilities, was illegally accessed by hackers during testing, highlighting AI security challenges and putting the safety-focused company in a difficult position.....

Microsoft has launched a one-time voluntary retirement program for some employees in the United States, covering those with many years of service and an age-plus-years-of-service total of 70 or more, affecting about 8,750 people (7% of U.S. employees). This move is related to the development of artificial intelligence and the wave of tech layoffs. Eligible employees and managers will be notified on May 7th.

AI unicorn DeepSeek has launched its first external fundraising round, targeting a valuation over $20 billion, with Tencent and Alibaba expressing interest. Despite relying on founder Liang Wenfeng's quantitative trading funds and having no urgent need for capital, the move aims to retain key talent through equity incentives amid recent departures of core members like R1 paper author Guo Daya and researcher Wang Bingxuan.....

NVIDIA CEO Huang Renxun wrote a letter to OpenAI CEO Altman, announcing that the company's thousands of employees have already started using the Codex programming tool based on GPT-5.5, covering multiple departments such as engineering, legal, and marketing, and achieving significant results.

At the Microsoft AI Tour annual event, Doushen Education launched the new 'Doushen AI Short Drama Platform,' built on a multimodal AI architecture integrating text comprehension, image generation, video generation, and intelligent dubbing. It covers scriptwriting, storyboard breakdown, and character setting, marking a major breakthrough in AI-driven content creation.....

Tencent released and open-sourced the new AI model 'Hunyuan Hy3 Preview', the most intelligent in its series. Upgrades cover complex reasoning, instruction following, contextual learning, code processing, and agents. It uses a hybrid expert architecture combining fast and slow thinking, with 295 billion parameters, to enhance overall performance and intelligence.....

Meta launches the Model Capability Initiative (MCI), leveraging employee work behavior data to accelerate large language model training. CEO Zuckerberg and Chief AI Officer announce ongoing improvements to the new model 'Muse Spark', but to speed up progress, daily work data will be integrated.....

Ping An Technology's 'Medical Large Model 3.5' achieved a score of 57.27 in a global medical AI evaluation, surpassing Meta and OpenAI to rank first. The test, involving 262 doctors from 60 countries and 5,000 high-fidelity dialogues, highlights the model's excellence in complex medical scenarios, showcasing Ping An's leadership in medical AI.....

On April 23, Huawei released the new HarmonySpace 6 cockpit, upgrading in-cabin perception, smart interaction, and mobile audio-visual functions. The next-gen assistant Xiaoyi, powered by MoLA 2.0 with billions of parameters, uses multimodal AI for full-scenario conversations, handling navigation and control tasks simultaneously, enhancing in-car intelligence.....

Meta has upgraded its 'Account Center' to 'Meta Account,' enabling unified digital identity management across Facebook, Instagram, Threads, and hardware like AI glasses and Quest headsets. Users can centrally manage all linked accounts via a single dashboard, signaling Meta's push to break down barriers between its apps and devices, fostering a highly integrated ecosystem.....

Singapore is transforming from a traditional East-West gateway into a 'neutral zone' for AI, attracting global top firms to avoid geopolitical risks. For Chinese startups, it offers access to international markets with reduced political bias; for US companies, it serves as a key base to bypass talent restrictions and attract global talent, highlighting its strategic hub role in tech competition.....

Meta announced the layoff of about 8,000 employees (10% of its workforce) to cope with the high infrastructure and talent costs brought by the transition to artificial intelligence. The move reflects a widespread anxiety in the technology industry: while AI improves efficiency, the expenditure on hardware and top talent is significant, forcing companies to streamline their organizations to maintain profitability. The CEO of Meta had previously warned investors about the pain of the transition.

Huawei launches HarmonySpace 6, with three upgrades: in-cabin sensing, smart interaction, and mobile audio-visual. The cabin agent Xiaoyi uses MoLA 2.0 architecture with a multimodal model of hundreds of billions of parameters, enhancing understanding, decision-making, and execution as a full-scenario AI assistant for navigation, vehicle control, chat, and lifestyle services.....

Google open-sourced Stitch's core format, DESIGN.md, a machine-readable protocol addressing brand consistency in AI Agent UI generation. It integrates YAML design tokens (e.g., colors, fonts) with plain-text annotations, providing AI with intuitive design guidelines and logical support to generate brand-aligned UIs while adhering to WCAG accessibility rules.....

On April 24, 2026, OpenAI released its most powerful AI model, GPT-5.5, achieving a qualitative leap in intelligence and significantly improving energy efficiency. This model is designed for the era of intelligent agents, capable of autonomously planning, calling tools, and executing complex multi-step tasks. It performs exceptionally well in code generation, scientific research, and enterprise automation. The core breakthrough is shifting from passive response to active execution, significantly enhancing autonomy.