Yandex has published an official statement refuting rumors that its neural networks (YandexGPT, YandexART, and others) are based on the open-source Chinese model Qwen. The company claims that it uses its own architectural solutions and trains models on its own data.
Яндекс
Elon Musk announced that SpaceX engineering data will become part of the training set for the next version of xAI's Grok AI model. According to him, this will allow the model to better understand complex technical and engineering concepts.
xAI
SpaceX
Chinese tech giant Tencent has merged its large language model team Hunyuan with the multimodal division. The new combined unit will be led by Yao Shunyu.
Tencent
At the WAIC forum, a discussion was held with the participation of six companies working on world models. Technical leaders diverged on key issues: whether world models should aim for physical accuracy or plausibility, which capabilities are more important, and what will become the first consensus in the industry. The transition from demonstrations to real-world deployment was also discussed.
According to Google, artificial intelligence is used in 68% of professions, but it does not replace entire professions, only specific tasks. The technology is not yet capable of fully displacing humans.
Google/DeepMind
AMD announced the X100 series of embedded processors based on the Zen 5 architecture with RDNA 3.5 graphics and an XDNA 2 neural processing unit for robotics. The chips are designed for a 10-year lifespan, have a TDP range of 45 to 120 W, and an operating temperature range of -40 to 105 °C. The flagship X199 model with 16 cores and 40 compute units demonstrated superiority over the Intel Core Ultra in tests, though the comparison was qualified.
Intel Corporation
NVIDIA
Due to the AI boom, American tech giants are increasingly borrowing funds for data center construction, but lenders are raising interest rates. Meta recently secured financing at a rate above 7%, which is 0.4 percentage points higher than last year, indicating growing lender concerns about the profitability of such projects.
Meta
Anthropic
Broadcom
At the 2026 Automotive Thermal Systems Conference, company president and Tsinghua University Advanced Battery Research Institute director Li Yantao announced the world's first bionic intelligent battery system. According to project data, the system reduces the development cycle by over 60%, lowers research costs by more than 50%, improves battery temperature uniformity by 31%, and reduces peak cell temperature by 8 degrees Celsius.
The stable version of Android Studio Quail 2 has been released, enhancing the capabilities of the Gemini AI agent: now multiple dialogs can be run in parallel. The IDE integrates the LeakCanary tool for detecting memory leaks, and crash analysis has been improved. Additionally, the Studio Labs feature for testing experimental AI functions has been stabilized.
Google/DeepMind
At WAIC 2026, experts noted a paradigm shift: the AI industry is moving from demonstrating model capabilities to strict financial accounting ('settling accounts'). Three key cost areas are discussed: infrastructure (total investments exceeded $700 billion with revenue of only $50 billion), inference (cost per token dropped 8–10 times, but total costs rose), and capital (investors demand real economics). Models remain important, but the focus shifts to product, agent capabilities, infrastructure, and trusted data.
Anthropic
Moonshot AI
German Gref stated in the Federation Council that at Sber, artificial intelligence makes all decisions regarding individuals. This applies to both credit and non-financial decisions.
Сбер
Сбербанк
Google DeepMind has announced Gemini 3.5 Flash Cyber, a specialized model based on 3.5 Flash, fine-tuned for vulnerability search, verification, and fixing. The model will be available through a limited pilot program to governments and trusted partners for proactive mitigation of critical vulnerabilities.
Google/DeepMind
Google DeepMind
NVIDIA introduced the Vera Rubin platform, designed for continuous post-training of agentic AI models. The key metric becomes 'intelligence per dollar', combining token cost and training efficiency. The platform promises a fourfold reduction in GPU count compared to Blackwell.
NVIDIA
Prime Intellect
Perplexity AI
Together AI
According to The New York Times, Meta Platforms is discussing the lease of computing capacity from AI company Anthropic. The potential deal could be worth $10 billion.
Meta
Anthropic
Work management platform Smartsheet has built a remote MCP server on AWS, providing AI agents with structured access to data and functions. The architecture includes AWS Fargate, Amazon Bedrock, and Neptune, supporting both internal AI assistants (Smart Assist) and external clients (Amazon Quick, Claude Desktop). Optimizations have saved over 3 billion tokens, and an access management system is implemented for enterprise customers.
Smartsheet
Amazon Web Services
Anthropic
Amazon Quick is an AI assistant for sales teams that automates administration, lead prioritization, meeting preparation, and CRM updates. The tool integrates with Salesforce, Outlook, and other systems, allowing sales reps to focus on selling.
Amazon Web Services
Chinese startup Moonshot AI has unveiled a new language model, Kimi K3, claiming it is on par with top US counterparts such as GPT-4o and Claude 3.5 Sonnet. The model achieves high scores in tests and is available in the Kimi app.
Moonshot AI
Analysts at Goldman Sachs noted that Moonshot AI's Kimi K3 model has reached the key level of intelligence required for global deployment of open-source AI models. Meanwhile, competition among high-end models is becoming increasingly fierce.
Moonshot AI
Anthropic is including the Fable 5 model in Max and Team Premium subscriptions on a permanent basis, starting July 20, 2026, with 50% limits. Pro and Team Standard will receive a one-time credit of $100. Competition from GPT-5.6 Sol and Kimi K3 forced Anthropic to abandon plans to make Fable 5 available only via API.
Anthropic
OpenAI
Moonshot AI
Chinese company Moonshot AI has released a large language model, Kimi K3, capable of completing the full chip design cycle in 48 hours. The development has received approval from Elon Musk and demonstrates the potential of AI in automating complex engineering tasks.
Moonshot AI
At the WAIC conference, Baidu Intelligent Cloud introduced the Tianchi supernode, designed to push the boundaries of scaling domestic computing power. The new product is aimed at supporting large-scale AI workloads.
Baidu
Baidu has unveiled its new AI chip, the Kunlunxin M100, which is entirely developed in China and positioned as an analogue to the NVIDIA H20. The chip is designed for high-performance computing and training AI models.
Baidu
Chinese AI startup Moonshot AI (Yue Zhi An Mian) reported a shortage of computing resources after releasing its new K3 model. As a result, the company has temporarily suspended registration for new users in the consumer segment (C-end) to balance load and ensure service stability.
Moonshot AI
The MERA Reason leaderboard has been launched to evaluate reasoning models in Russian. It includes four task sets: Luzitania (olympiad-level mathematics), TMath (a wide range of problems), MMReD (dense context reasoning), and an upcoming component. The goal is to objectively measure a model's ability to build chains of reasoning, not just reproduce knowledge.
Т-Банк
Alibaba/Qwen
Prompt injection is a structural defect in large language models where untrusted text is executed as an instruction. Filters and system prompts are ineffective: even the best defenses are breached in 90% of cases under adaptive attacks. Working approaches—Dual-LLM, CaMeL, and the Two Meta rule—build isolation rather than trying to distinguish good from evil.
OpenAI
Anthropic
Google DeepMind
Microsoft