SoftBank Group announced an investment of up to €75 billion to build AI data centers with a total capacity of 5 gigawatts in France, marking its largest AI infrastructure investment in Europe. The first phase involves €45 billion to develop a 3.1GW data center cluster in Hauts-de-France, located in Dunkirk, Beaucamps-Ligny, and Bouchain, expected to be completed by 2031. SoftBank will also collaborate with Schneider Electric to establish an advanced manufacturing hub at the Port of Dunkirk for producing modular data center units and racks, advancing local supply chain development. The project is projected to create thousands of high-skill jobs, leveraging France's low-carbon electricity. SoftBank has formed a joint venture with Sesterce to develop a 1GW campus in Beaucamps-Ligny and launched a €100 million fund to support local AI applications.

DeepSeek has introduced the 'Thinking with Visual Primitives' multimodal framework, embedding bounding boxes and point coordinates into reasoning chains to solve the 'Reference Gap' issue caused by language ambiguity in current multimodal models. Built on the DeepSeek-V4-Flash MoE language model and a proprietary ViT vision encoder, it achieves a visual token compression ratio of up to 7056x, requiring only about 90 KV cache entries to process an 800×800 image. The team curated over 40 million cold-start training samples, and the model significantly outperforms others in topological reasoning tasks such as maze navigation (66.9% vs GPT-5.4’s 50.6%) and path tracking (56.7% vs 46.5%). The paper emphasizes that 'pointing more accurately' matters more than 'seeing more'.

AI large model company MiniMax signed a counseling agreement with CITIC Securities on May 29 to officially launch its A-share IPO process, aiming to achieve an A+H dual listing on the STAR Market. Founded in early 2022, MiniMax set a record in January 2026 by becoming the fastest AI company to go public on the Hong Kong Stock Exchange. Its share price has risen over 400% from its offering price, reaching a market cap of approximately HK$263.4 billion. Business metrics are strong: around 300 million global users, serving over one million customers, with ARR more than doubling in the past two months to exceed $300 million. Previously, Zhipu AI also engaged Guotai Junan and Haitong Securities to advance its STAR Market listing, potentially bringing two major Chinese AI firms together on the A-share market.

NVIDIA and Tsinghua University, along with other institutions, released Gamma-World, the first framework to systematically address identity symmetry and computational efficiency bottlenecks in multi-agent world models. By encoding player identities using vertices of a standard simplex, it supports arbitrary numbers of agents without retraining; a hub token mechanism reduces computational complexity from quadratic to linear. The model excels in collaborative two-player Minecraft tasks and demonstrates zero-shot generalization to four-player scenarios using only two-agent training data. A three-stage distillation approach enables real-time inference at 24 FPS. This framework provides scalable interactive data generation for Physical AI, potentially overcoming the scarcity of physical interaction datasets.

Apple has introduced the PICO perceptual image codec, the first practical learned codec systematically optimized for human visual perception. Unlike traditional codecs that optimize mathematical metrics like PSNR, PICO directly optimizes for human subjective quality. Key innovations include a single-pass context model resolving entropy coding speed bottlenecks, TextFidelityLoss to prevent GAN hallucinations in text regions, and TilingArtifactLoss to eliminate tiling boundary artifacts. Large-scale human subjective evaluations show PICO reduces bitrates by 57%-70% compared to AV1, VVC, and JPEG AI while achieving millisecond-level encoding and decoding speeds on iPhone devices.

Kuaishou has open-sourced the Keye-VL-2.0-30B-A3B multimodal large model, the first to introduce DeepSeek’s DSA sparse attention to multimodal scenarios. With 30 billion total parameters and 3 billion activated parameters, it outperforms larger models like Qwen3-VL-235B-A22B across multiple video understanding benchmarks including VideoMME V2 and LongVideoBench. DSA reduces long-sequence prefill costs by 50%, enabling efficient understanding of hour-long videos. The model also natively integrates agent collaboration mechanisms in a multimodal base, leading in benchmarks such as LivecodeBench and SWE-bench. It employs MOPD to mitigate catastrophic forgetting in multi-task learning and Context-RL to suppress hallucinations.

Waymo’s sixth-generation robotaxi, Ojai, has begun trial operations in three cities including San Francisco. The new model reduces sensor count by 42% and cuts per-unit cost by $75,000, accelerating the commercial scale-up of autonomous driving. Despite fewer sensors, safety is not compromised—enhanced sensor fusion and software algorithms deliver equal or better perception performance. This marks a significant step for Waymo in lowering hardware costs and expanding operational scale, paving the way for broader deployment of driverless taxi services.

The unreleased GPT-5.6 model, codenamed "Iris," has been spotted in system logs with a context window of 1.5 million tokens, expected for release in early June. This leak follows the earlier accidental exposure of GPT-5.5 on the Codex platform, drawing renewed attention. A 1.5 million-token context will greatly enhance the model’s ability to handle long documents and large codebases. OpenAI is currently in the confidential IPO filing stage, targeting a valuation exceeding $850 billion, with the new model’s rollout closely tied to its上市 timeline.

Arm has open-sourced the Metis security framework, which uses agentic AI and semantic reasoning to detect complex cross-component vulnerabilities. In internal benchmarks using GPT-5.5-Cyber as the base model, Metis achieved a 98% vulnerability detection accuracy rate, compared to just 6% for traditional SAST tools—representing a tenfold improvement in true positive rate and roughly 50% fewer false positives. Released under the Apache 2.0 license, Metis supports any OpenAI-compatible LLM and covers C, C++, Python, Go, TypeScript, and Rust. It can integrate with existing SAST tools to validate findings and provides actionable vulnerability explanations via natural language summaries to accelerate remediation.

Boston Children's Hospital reported that deploying an enterprise-grade AI system has enabled diagnosis of 40 rare diseases, saving 60,000 work hours and reallocating $7 million in labor costs. Over one-third of staff now use AI daily, automating more than 50 workflows across research, clinical, and administrative operations. The hospital established a secure internal ChatGPT environment to ensure data privacy. This case provides concrete evidence of AI scalability in healthcare, demonstrating impact across three key dimensions: number of diagnoses, time savings, and cost optimization.