DeepSeek founder Liang Wenfeng stated in a new round of financing totaling 70 billion RMB that the core goal is advancing artificial general intelligence (AGI), rather than short-term commercialization. Potential investors include the National AI Industry Investment Fund (planning to invest around 10 billion RMB), Tencent, IDG Capital, and Monolith Capital, with valuation potentially reaching $45 billion. The funding round is in its final stage; if completed, it will set a record for the largest initial financing by a Chinese tech startup. DeepSeek will continue its open-source AI model strategy while expanding into agent-based AI, driving Asia's AI models toward an industrial model based on compute tokens.

Anthropic released an update on Project Glasswing's first month, using the Claude Mythos Preview model to identify over 10,000 high or critical severity vulnerabilities in key software. Partners Cloudflare and Mozilla reported more than a 10-fold increase in vulnerability discovery rates. Over 1,000 open-source projects were scanned, estimating 6,202 high-risk vulnerabilities found, including a certificate forgery flaw in the wolfSSL crypto library. Anthropic noted that the bottleneck in software security has shifted from discovery to validation, disclosure, and patching. The company also launched Claude Security enterprise services, a web verification program, and opened access to relevant tools for qualified security teams. Japan's Ministry of Health, Labour and Welfare and the Minister of Finance held emergency meetings due to Mythos' vulnerability detection capabilities, discussing cybersecurity measures for healthcare and financial systems.

On May 22, Zhipu AI launched the GLM-5.1 High-Speed API, now available to select enterprise customers. The model achieves an output speed of 400 tokens per second, setting a new global record for large model API speeds. This version maintains the flagship-level intelligence of GLM-5.1 while delivering ultra-fast response times, challenging the industry assumption that 'fast means small.' The core technology is powered by the TileRT inference engine, which uses static orchestration to compile the entire computation graph into a persistent Engine Kernel resident on GPU, refining scheduling units from operator level to Tile-level microtasks, significantly reducing scheduling and GPU memory access overhead. Real-world testing shows code generation efficiency improved approximately 10x, making it ideal for AI programming, real-time interaction, and business decision-making scenarios.

Microsoft Research Frontier AI Lab released the Fara1.5 series of browser computer-use agent models, available in 4B, 9B, and 27B variants, built upon Qwen3.5. On the Online-Mind2Web benchmark, Fara1.5-27B achieved a 72% task success rate, significantly surpassing OpenAI Operator (58.3%) and Gemini 2.5 Computer Use (57.3%). The model employs an observe-think-act loop mechanism, trained on approximately 2 million samples. Its synthetic data pipeline, FaraGen1.5, created six verifiable simulation environments to generate high-quality training trajectories. The model pauses and requests user confirmation when handling personal information, ambiguous tasks, or irreversible actions, ensuring safety and controllability.

On May 22, 2026, Li Chao, spokesperson for China's National Development and Reform Commission (NDRC), revealed at a press conference that the commission is preparing supporting documents to accelerate the implementation of 'AI+', along with conducting research on AI legislation. The new policies will encourage central state-owned enterprises to open high-value application scenarios and develop industry-leading AI applications. Additionally, efforts will guide domestic large language models to accelerate compatibility with domestically produced AI chips, promoting supply chain autonomy and control. The quota for re-lending programs supporting technological innovation and upgrades will be expanded from 500 billion to 1.2 trillion RMB, with interest rates reduced from 1.75% to 1.25%, and coverage extended to 14 key areas including AI. Li reiterated that China has never required tech companies to reject foreign investment and remains committed to its fundamental policy of opening up.

On May 22, OpenAI announced the opening of its first applied AI lab outside the U.S., located in Singapore, as part of the 'OpenAI for Singapore' initiative, committing over 300 million SGD in investment. The lab will create over 200 local technical jobs in the coming years, focusing on deploying AI in public services, finance, and digital infrastructure. OpenAI Academy will also establish a Singapore branch and host teacher programming hackathons. Concurrently, Singapore's IMDA updated its agentic AI governance framework based on feedback from over 60 companies, adding guidance on multi-agent systems, third-party agents, and automated bias, incorporating real-world cases such as Tencent's CodeBuddy.

Modal Labs, an AI serverless cloud computing company, completed a $355 million Series C funding round led by General Catalyst and Redpoint, achieving a post-money valuation of $4.65 billion. Since September 2023, the company has grown fivefold, with annualized revenue exceeding $300 million. Modal specializes in designing cloud infrastructure for AI workloads, supporting elastic inference, reinforcement learning, and large-scale batch processing, featuring capabilities such as GPU snapshots for fast cold starts and scaling from zero to 1,000 GPUs within minutes. Its sandbox feature, launched in 2023, has run over one billion times and now accounts for more than one-third of total revenue. New investors include Menlo, Bain Capital Ventures, and Accel.

AI coding tool company Cursor has reached an annualized revenue of $3 billion, serving over 3,000 enterprise customers paying more than $100,000 annually, making it one of the fastest-growing startups in history. Meanwhile, Cursor released an SDK enabling developers to build custom agents using Composer 2.5, offering both Python and TypeScript versions. Notably, SpaceX holds the right to acquire Cursor at a $60 billion valuation within a 30-day window after Cursor goes public. On Microsoft's side, due to the high token-cost of Claude Code, the company plans to require nearly 100,000 engineers to discontinue using Claude Code by the end of June, switching instead to its own GitHub Copilot CLI.

NVIDIA introduced Nemotron-Labs-Diffusion, an 8B-parameter language model supporting autoregressive, diffusion, and speculative decoding modes. In diffusion mode, its decoding efficiency reaches 6x that of Qwen3-8B, significantly accelerating inference while maintaining generation quality. The model demonstrates the feasibility of converting traditional autoregressive models to support diffusion decoding, opening a new path for LLM inference acceleration. Additionally, ByteDance released the open-source multimodal model Lance on the same day, featuring only 3 billion active parameters, trained from scratch on 128 A100 GPUs, supporting image and video understanding, generation, and editing under the Apache 2.0 license.