At WWDC 2026, Apple officially launched its long-anticipated AI-enhanced Siri, upgrading it from a voice assistant to a full-featured conversational AI companion, and introduced a standalone Siri app. The new Siri is integrated into the iPhone's Dynamic Island and features screen awareness, context understanding, and personalized writing assistance (「Write with Siri」), capable of executing complex cross-app tasks. The underlying architecture is based on Apple Foundation Models co-developed with Google, incorporating Gemini technology, and supports both on-device and Private Cloud Compute execution. The new features will roll out in beta later this year, initially available only in English, with no launch planned for the EU or mainland China.

On June 8, OpenAI announced it has confidentially submitted an S-1 registration draft to the U.S. Securities and Exchange Commission, with a company valuation exceeding $850 billion. Goldman Sachs and Morgan Stanley are serving as lead underwriters. This follows a similar move by Anthropic (valued at $965 billion) just a week earlier, marking another major AI firm advancing toward public listing. Combined with SpaceX’s Nasdaq debut on June 12 (valued at $1.77 trillion), these three AI super unicorns now have a total valuation exceeding $4 trillion and are expected to raise over $150 billion collectively. OpenAI stated that the exact上市 timeline remains undetermined and may include an employee stock repurchase program to provide liquidity ahead of上市.

On June 8, NVIDIA unveiled the open-source physical AI foundation model Cosmos 3, built on a hybrid Transformer architecture integrating visual reasoning, environment generation, and action prediction. It supports multimodal processing of text, images, video, environmental sounds, and physically accurate motion trajectories, reducing physical AI training and evaluation cycles from months to days. NVIDIA also announced the formation of the Cosmos Consortium, with founding members including Agile Robots, Black Forest Labs, and Runway. The Cosmos 3 series includes Super, Nano, and an upcoming Edge version. It has ranked first in multiple physical AI benchmarks and is available to developers via build.nvidia.com, Hugging Face, and GitHub.

Xiaomi's MiMo team, in collaboration with the TileRT systems group, launched MiMo-V2.5-Pro-UltraSpeed, becoming the first system to achieve over 1,000 tokens per second decoding speed on a trillion-parameter language model, with peak performance nearing 1,200 tokens/sec, running on standard 8-GPU consumer-grade nodes without custom chips. Key technologies include FP4 (MXFP4) quantization applied exclusively to MoE expert modules to reduce memory bandwidth pressure; DFlash parallel speculative decoding with an average acceptance length of 6.30; and optimizations in the TileRT runtime system using persistent kernels and Warp specialization. API trials run from June 9 to 23, priced at three times the standard version but delivering approximately tenfold speed improvements. Partial weights have been open-sourced.

UIUC, UC Berkeley, and Chroma jointly released Harness-1, an open-source search agent based on the OpenAI gpt-oss-20B architecture (20 billion parameters), achieving a 73% information recall accuracy—outperforming GPT-5.4 (70.9%) and Tongyi DeepResearch 30B. The key breakthrough is the 「state externalization」 mechanism, which moves bookmarks, evidence management, and verification records from the model’s internal state to an external structured environment, significantly reducing cognitive load. Training used only 899 supervised fine-tuning trajectories and 3,453 reinforcement learning queries, demonstrating far superior data efficiency compared to peers. Released under the Apache 2.0 license for commercial use, it represents a shift in AI agent development from scaling model size to optimizing execution environments.

Generalist AI has completed a $400 million Series B round led by Radical Ventures, with participation from NVIDIA, Bezos Expeditions, and Fei-Fei Li. Total funding now exceeds $500 million. The company is developing a general-purpose AI brain 「GEN-1」 designed to work with any robotic hardware, achieving 99% task reliability—far surpassing the industry benchmark of 64%. Its technology is based on large-scale data collected from human operations, validating the 「scaling laws」 in physical AI. The founding team hails from Google DeepMind and Boston Dynamics, positioning the company as 「intelligence infrastructure」 for the robotics era, competing with vertically integrated players like Figure AI.

An Anthropic report reveals that over 80% of the codebase for Claude is now self-generated by the model—up dramatically from single-digit percentages in early 2025. The company believes AI technology has entered stage four of five, characterized by autonomous code execution and task delegation. The fifth stage will involve AI autonomously designing and training the next generation of models. Import AI also notes that Anthropic’s codebase merge volume in 2026 has grown eightfold compared to 2024, showing early signs of 「Recursive Self-Improvement」 (RSI). Despite repeated global calls to pause AI development, Gartner forecasts global AI spending will reach $2.59 trillion in 2026, a 47% year-on-year increase.

According to OpenRouter’s latest data, global LLM inference volume reached 36.1 trillion tokens for the week of June 1–7, 2026, a 13.5% increase from the previous week. China recorded 14.19 trillion tokens, up 27.49% week-on-week, marking the sixth consecutive week surpassing the US (3.2 trillion tokens, down 24.53%). Among the top five most-used models globally, Chinese models occupy the top four spots: DeepSeek-V4-Flash remained first for the third consecutive week with 3.69 trillion tokens, Tencent Hy3 Preview ranked second with 2.94 trillion, newly launched MiniMax M3 entered third with 2.5 trillion in its first week, and Xiaomi MiMo-V2.5 came fourth with 2.19 trillion. DeepSeek has led all vendors for four straight weeks with 6.75 trillion tokens.

Microsoft Research introduced Lens, an efficient text-to-image generation model with only 3.8 billion parameters, yet outperforming larger counterparts such as Hunyuan-Image-3.0 (80 billion parameters). The key lies in the high-quality Lens-800M dataset—800 million image-text pairs generated by GPT-4.1, with each caption averaging around 100 words, vastly superior to short web-scraped alt-text. Training compute cost is only one-fifth of comparable models. Lens uses FLUX.2 semantic VAE and GPT-OSS text encoders, supports multilingual input, and is optimized via reinforcement learning. The lightweight Lens-Turbo can generate images in four steps, taking less than one second on an H100. Code and weights are open-sourced under the MIT license, though limited to research use only.

London-based startup PhysicsX has closed a $300 million Series C round, reaching a $2.4 billion valuation—more than double its nearly $1 billion valuation a year ago—with total funding approaching $500 million. The round was led by Singapore’s sovereign wealth fund Temasek, with participation from NVIDIA, Atomico, and Siemens. Founded by two former F1 engineers, the company leverages AI to compress complex design and simulation workflows in aerospace and semiconductor industries—from months down to seconds. Funds will support platform development, AI research, U.S. market expansion, and a new office in Singapore. With about 350 employees, PhysicsX is emerging as a flagship AI-native engineering software company.