Skip to content
Source · Daily Brief

AI Morning Brief — 7 July 2026

Tencent released Hy3, a 295-billion-parameter open-source model under Apache 2.0 licensing that claims to match models two to five times its size, while Anthropic faced criticism overnight for reportedly tracking users without disclosure even as it published notable new interpretability research.

Top stories

  • Tencent drops Hy3. The 295B mixture-of-experts model activates 21B parameters at inference time and reports a 5.4 percent hallucination rate, down from earlier Hy releases. Tencent is offering a free API on OpenRouter for two weeks and has published weights under Apache 2.0. via THE DECODER; via Simon Willison
  • Anthropic’s global-workspace paper. Anthropic posted research showing language models contain a “J-space”—a verbalizable working memory that sustains multi-step reasoning; removing it leaves factual recall intact while degrading complex problem-solving, a pattern the authors liken to deliberate versus automatic processing in humans. via Hacker News
  • Anthropic outed for secret Claude tracker. Ars Technica reported that Anthropic ran an undisclosed monitoring system tracking Claude usage in China, a disclosure that drew community criticism and sparked commentary about the lab’s standing with users. The company reportedly ended the experiment once it was publicised. via Ars Technica – AI
  • First AI-executed ransomware, with caveats. The JADEPUFFER operation deployed a language model to autonomously execute an intrusion—stealing credentials and destroying databases—but new reporting clarifies that a human selected the target, built the infrastructure, and supplied credentials. The attack marks a milestone in agentic security risk rather than fully autonomous cybercrime. via TechCrunch – AI; via THE DECODER
  • Nvidia’s Kyber NVL144 slips to 2028. SemiAnalysis reported the next-generation AI server rack has been pushed back more than a year owing to circuit board manufacturing problems; the Rubin Ultra variant has also been canceled. Asian suppliers fell by double digits on the news. via THE DECODER
  • China ends humanlike AI personas. New Beijing regulations require ByteDance and Alibaba to shut down features that let users build and chat with custom AI companions before July 15. The rules target anthropomorphic chatbot design more broadly, affecting how both Doubao and Qwen surface to Chinese users. via THE DECODER

Who shipped

Tencent had the most significant product delivery of the cycle with Hy3, switching from a restrictive community license to Apache 2.0 and pairing the release with a free two-week API window. Anthropic published its global-workspace interpretability paper, advancing its mechanistic interpretability research agenda, though the tracker disclosure dominated press coverage. Sakana AI launched Sakana Translate, a Japanese-English-Chinese translation tool built on the Namazu model series, shipping translate, proofread, and ask modes together.

Open-source pulse

Hy3 (295B total, 21B active, Apache 2.0) was the headline release. Sberbank subsidiary ai-sage separately posted GigaChat3.5-432B-A28B with same-day GGUF support. In the community, Kyutai’s Pocket TTS earned attention for cloning a voice from five seconds of audio and running entirely on CPU under an MIT license.

Money, infra & hardware

SK Hynix is set to launch a multibillion-dollar US IPO on Friday, adding a second AI memory maker to US markets alongside its domestic rival. Microsoft cut approximately 4,800 positions—2.1 percent of its workforce—with commercial sales and Xbox absorbing the largest share. Amazon announced it is closing Mechanical Turk to new customers on July 30, retiring a two-decade-old crowdwork platform that helped label data for some of the earliest AI training pipelines.

Quiet corners

OpenAI, Meta, and Mistral were absent from primary-source coverage during the window; xAI was similarly quiet—notable given that DeepSeek V4 and Nvidia’s SuperPod 950 compute cluster are both reported to be imminent.

By the numbers

  • 319 stories in 24 h across 23 sources
  • Most-mentioned model: DeepSeek
  • Most-mentioned lab: Anthropic
  • Notable absences: OpenAI, Meta, Mistral, xAI

Compiled by AI Feed’s editor from all 319 headlines published on 7 July 2026.