According to multiple independent evaluations, Zhipu AI's open-source model GLM-5.2 has become the first open-source model recognized by the industry as reaching frontier-level performance. It leads on the Artificial Analysis AA-Briefcase benchmark and has received spontaneous praise from developers including Jeremy Howard. Citing data from Tsinghua University's Tang Jie team, GLM-5.2 scored 42.8 on the BridgeBench reasoning benchmark, surpassing Fable 5, with a cost approximately 1/10 that of closed-source models and a speed of up to 300 tok/s. The model is released under the MIT license and can run entirely locally. Analysts note that freely distributable open weights rendered U.S. export controls on Anthropic's Fable 5 ineffective within 48 hours of GLM-5.2's release, confirming the view that 「export controls cannot block open source」.

OpenAI, preparing for IPO, has announced the recruitment of Noam Shazeer. Shazeer is a primary author of the foundational paper 「Attention Is All You Need」 and a key contributor to the Transformer architecture, previously serving as co-lead for Gemini at Google. This move continues the trend of top AI talent flowing between Anthropic and OpenAI, widely seen as a significant indicator of talent competition. Shazeer's deep expertise in Transformer architecture brings direct value to OpenAI's cutting-edge model development and adds technical credibility to its IPO narrative.

AI imaging company Midjourney has announced its entry into personal health with the launch of the Midjourney Scanner, a full-body ultrasound device. Using a ring array of underwater ultrasonic sensors, the device generates high-precision 3D body models in just 60 seconds—far faster than traditional MRI—and does not itself use AI technology. The company plans to deploy 50,000 scanners, with the first unit set to debut at the 「Midjourney Spa」 in San Francisco in 2027. Midjourney also disclosed it is advancing four hardware and four software projects, with at least two new hardware products expected soon, marking its aggressive shift from generative AI to medical hardware.

A study published in the AI journal of The New England Journal of Medicine shows that OpenAI's o3 model played a role in diagnosing pediatric rare genetic diseases, helping identify 18 new diagnoses among 376 previously undiagnosed cases. OpenAI President Greg Brockman emphasized the model has been available for over a year, suggesting even greater potential with updated models. Meanwhile, GPT-5.5 Instant has demonstrated performance on par with frontier models in health-related Q&A across hundreds of doctors in 60 countries, indicating OpenAI's focused push into vertical AI applications in healthcare.

According to TLDR, OpenAI plans to release GPT-5.6 next week, potentially including Mini and Pro versions. Key upgrades include a 1.5 million token context window, enhanced long-horizon coding capabilities, and faster Codex response speeds, along with a competitive pricing strategy. This update focuses on extending context length and practical utility in programming scenarios, aiming to solidify OpenAI's competitive edge against Anthropic in the AI coding market. Specific performance metrics and pricing details have not yet been disclosed.

AI data platform Databricks plans to acquire AI-powered security operations (SOC) platform Panther to strengthen its security software strategy. Concurrently, Amazon AWS launched the Continuum security service, capable of rapidly identifying, prioritizing, validating, and fixing code vulnerabilities. These moves reflect how AI companies are accelerating their expansion into security operations. Additionally, HPE noted that as GPU clusters and high-throughput AI traffic grow, network connectivity is emerging as the third major bottleneck in AI data centers after space and power, highlighting new infrastructure challenges.

Security researchers have discovered a large-scale malware distribution campaign on GitHub involving approximately 10,000 repositories disguised as legitimate projects. While appearing to be normal open-source initiatives, these repos are actually used to distribute malicious software, posing supply chain security risks to developers. This incident underscores the growing importance of code provenance verification and supply chain security amid the rising prevalence of AI-assisted programming and open-source dependencies.

Top content creator PewDiePie has released Odysseus, a self-hosted AI workbench, which gained over 30,000 GitHub stars within three days. More than just a chat interface, the project is a full-featured workbench integrating local agents, email systems, Deep Research, document editor, and image editor—all running completely locally, positioning itself as an alternative to ChatGPT Plus and Claude Pro. Its built-in local agent can browse the web, edit files, execute code, and evolve through experience logging; the Cookbook feature automatically scans hardware, intelligently recommends, and deploys local models with one click. The surge in popularity reflects strong user demand for data sovereignty and decentralized AI.

Xiaomi has open-sourced Xiaomi Miloco 2.0, achieving 「Agentization」 of smart homes. Key upgrades include multimodal perception, proactive intelligence, continuous task handling, and household memory, making it the industry's first smart home solution to enable proactive AI services and family memory. The system processes camera and sensor data on-device, collaborates with cloud-based large models for decision-making, and ensures privacy by immediately discarding raw data, storing only locally, isolating it from the Agent, and auto-clearing after 30 days. Leveraging Mi Home's ecosystem of nearly 1.2 billion AIoT devices and MiMo's large model Agent capabilities, Xiaomi has built a hard-to-replicate moat.

World model company Jijia Vision has completed a 1 billion yuan Series B2 funding round, leveraging a 「dual pyramid」 system of data and algorithms to drive world models and embodied intelligence. Its GigaBrain-0 achieved a 51.67% success rate in RoboChallenge, outperforming Physical Intelligence's π0.5 by nearly 10 percentage points, while GigaWorld-Policy topped the RoboCasa365 leaderboard and defeated NVIDIA's GR00T N1.5. Commercialization is progressing on two fronts: the consumer robot 「ShiGuang S1」 has already secured hundreds of orders, and industrial deployments aim to reach 1,000 units within three years. The company unveiled a 12-month roadmap, with GigaBrain-1 scheduled for release in Q3 2026 and GigaBrain-3 trained on 1 million hours of video aiming to trigger emergent physical AGI capabilities.