Moonshot AI Releases Kimi K3: The World's First Open 3-Trillion-Class Model
Chinese AI lab Moonshot AI officially released Kimi K3 on July 16, making it the first open model in the 3-trillion-parameter class. With 2.8 trillion total parameters, a 1-million-token context window, and native vision capabilities, Kimi K3 is the largest open-weight model ever released — and it's competitive with the best proprietary systems available.
Built on a new-architecture Mixture-of-Experts design featuring Kimi Delta Attention (KDA), a hybrid linear attention mechanism, the model scored 1,687 on the GDPval-AA v2 benchmark. That places it third overall, behind only Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,748), but ahead of Claude Opus 4.8 (1,600). For an open-weight model, that's a remarkable achievement.
Kimi K3 launched with two variants: K3 Max for chat and agent tasks, and K3 Swarm Max for large-scale parallel processing. The full model weights are expected to be released by July 27 under a Modified MIT license, continuing the open-weight approach Moonshot AI established with the Kimi K2 family.
The release intensifies competition in the open model space and narrows the gap between open-weight and proprietary frontier models — a trend that has defined 2026's AI landscape.
Japan Commits $2.4 Billion to Build National AI Infrastructure for Robotics
Japan announced a massive national initiative on July 16 to purchase 27,500 next-generation NVIDIA Rubin GPUs along with 13,750 Vera CPUs to build what NVIDIA calls "the world's first national AI infrastructure." The project, led by Noetra Corp., has received ¥387.3 billion ($2.4 billion) in government funding.
The AI factory will provide the computing foundation for Japan's FRONTia Project — officially titled "Development of Multimodal Foundation Models with a View to AI Robotics and Physical AI" — launched by Japan's Ministry of Economy, Trade and Industry (METI). Dozens of major Japanese companies including Sony, SoftBank, Toyota-backed Preferred Networks, and NEC are collaborating on the effort.
The data center, expected to deliver 140 megawatts of capacity based on NVIDIA's DSX platform, is slated to go online in June 2028. Japan's bet on robotics AI reflects a strategic response to the country's aging population and labor shortages, positioning the nation to lead in physical AI and autonomous systems.
Five Eyes Agencies Warn AI Cyber Threats Are "Months, Not Years" Away
The Five Eyes intelligence alliance — the United States, United Kingdom, Canada, Australia, and New Zealand — issued a stark warning that frontier AI models will "fundamentally transform both offensive and defensive cyber capabilities," and that the timeline is "months, not years."
The joint advisory, published by CISA and its international counterparts, warns that advanced hacking capabilities provided by frontier models are expected to become broadly available within the year, despite efforts by AI companies to restrict access. The agencies specifically noted models like Anthropic's Fable 5 and OpenAI's Daybreak as examples of systems whose capabilities could be exploited.
The advisory listed five practical actions for organizations: reducing attack surfaces, patching more quickly, removing or isolating vulnerable legacy systems, improving identity management, and testing incident response procedures. The agencies emphasized that cybersecurity "is a core business risk and leadership responsibility."
On the positive side, the warning also acknowledged AI's defensive potential — organizations that integrate AI tools into security operations can detect vulnerabilities earlier, improve software quality, and respond faster to incidents.
Cloudflare Introduces Granular AI Bot Controls for Website Owners
Cloudflare launched a new granular AI bot management system that lets website owners separately control three categories of AI crawlers: Search, Agent, and Training bots. Starting September 15, 2026, new defaults will automatically block Agent and Training bots on ad-supported pages.
The move comes as AI agents increasingly crawl the web to complete tasks on behalf of users — a fundamentally different use case from traditional search indexing. Website owners can now make fine-grained decisions about which types of AI access they want to allow, reflecting growing concerns about AI training data scraping while still enabling beneficial AI-powered search and agent interactions.
This is one of the most concrete steps by a major infrastructure provider to give website operators meaningful control over how AI systems interact with their content — a question that has become increasingly urgent as agentic AI goes mainstream.
NVIDIA Open-Sources Diffusion Language Models With 2.4x Throughput Gains
NVIDIA released two open-weight diffusion language models that challenge the dominance of traditional autoregressive text generation. Nemotron-Labs-TwoTower, built on a frozen autoregressive backbone, achieves 2.42x higher throughput while retaining 98.7% of baseline quality — all without retraining.
The more advanced Nemotron-Labs-Diffusion model goes further with a tri-mode architecture that unifies autoregressive, diffusion, and self-speculation decoding in a single model. The 8B instruct variant produces six times more tokens per forward pass than a comparable autoregressive model and achieves 4x higher throughput on NVIDIA's serving benchmark when deployed with the SGLang inference framework on a GB200 GPU.
Both models are available as open weights under commercial licenses on Hugging Face. While these are research-scale models rather than frontier systems, they point toward a future where inference costs could drop dramatically — a development with major implications for making AI more accessible and affordable.