Daily AI RoundupsFriday, 28 August 2026

Tencent releases the 770B-parameter Hy4 preview with 49B active parameters and a 1M-token context window, while Google ships Gemini Omni 1.1 Flash and Anthropic opens its Model Hardware Standard research preview. Reuters reports Anthropic's MatX discussions and Nvidia's pause on some AI-cloud revenue-sharing deals; the edition also covers SK Hynix's $4B HBM plant, Meta and SoftBank financing, CXMT's 874% revenue growth, DeepSeek's capital search, Google-led cyber warnings and Gemini-based research.

• open-weight models
• multimodal AI
• AI hardware standards
• AI chips
• AI finance
• AI regulation
• AI security
• AI research

Stories in this roundup

1. 01
   Tencent releases the open-weight Hy4 preview, a 770B-parameter MoE with 49B active parameters and a 1M-token context windowTencent
2. 02
   Google releases Gemini Omni 1.1 Flash through the Gemini API with scene extension, camera controls and 4K upscalingGoogle
3. 03
   Anthropic opens the Model Hardware Standard research preview to labs and manufacturers for AI control of physical equipmentAnthropic
4. 04
   Anthropic discussed a $7B MatX acquisition and now explores a partnership to accelerate custom-chip development, Reuters reportsReuters
5. 05
   Nvidia pauses some revenue-sharing deals with AI cloud firms amid internal antitrust concerns, Reuters reportsReuters
6. 06
   US judge permanently blocks Trump administration restrictions on Anthropic, ruling they unlawfully punished protected speechCBS News
7. 07
   SK Hynix breaks ground on a $4B Indiana HBM packaging plant, targeting next-generation memory production in 2029SK Hynix
8. 08
   Meta projected it could spend up to US$10B a year on Anthropic's AI models, The Business Times reportsThe Business Times
9. 09
   SoftBank seeks another US$10B loan to refinance debt tied to its OpenAI stake, Bloomberg reportsBloomberg
10. 10
    CXMT reports first-half revenue rose 874% to 15B yuan, reversing a year-earlier loss as AI-driven DRAM demand surgedChosunBiz
11. 11
    DeepSeek seeks outside capital as rising model and compute costs test High-Flyer's funding, CNBC reportsCNBC
12. 12
    Google employees are testing Gemini 3.8 Flash Preview internally, Business Insider reportsBusiness Insider
13. 13
    Google, Microsoft, Anthropic and OpenAI join 100+ organizations warning conventional cyber defenses may not withstand rapidly improving AI attacksBBC
14. 14
    Google researchers report Gemini-based Co-Scientist closed the loop from hypotheses to real-world experiments across three scientific fieldsarXiv
15. 15
    Google DeepMind pilots a double-blind Gemini evaluation that keeps model weights and external test prompts privateGoogle DeepMind
16. 16
    llama.cpp adds native Qwen3.8-Flash-Next support for QSA sparse attention, MoE and PLE embeddingsllama.cpp