In the weeks following the launch of Chinese AI lab Moonshot AI’s Kimi K3 model, U.S. AI firms have scrambled to garner support for more open-weight models and released comparable offerings—even as their users continue to wait for a definitive breakthrough.
Meta
Meta CEO Mark Zuckerberg published a 6,500-word manifesto this week, outlining his vision of how to reach a “positive AI future.” In it, he declared that Meta was building personal superintelligence to improve everyone’s lives, and that every user would have an “exceptionally capable personal agent” to enhance their relationships, health, career, finances, home management, and hobbies.
“Everyone will have a personalized tutor and coach with a PhD in every subject and unlimited patience to help you learn anything you want,” he stated.
Doubling down on supporting more open models, Meta also announced the release of new models and access plans.
First, Mr. Zuckerberg announced that Meta planned to open the weights for its AI model Muse Spark 1.2. The social media giant also announced the release of Muse Glimmer. Meta described this as an open agentic model that runs on the user’s device, and said Muse Glimmer was trained on Muse Spark’s outputs using logit distillation, amongst other strategies.
“Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows. It’s small enough to run on a Mac or PC with a single consumer GPU, enabling use cases that range from local agents and function calling, to local coding, and LLM-as-a-judge evaluation,” announced the company.
For contrast, Moonshot’s open-weight Kimi K3 reported 2.8 trillion parameters.
Google saw a week of changes that involved both senior executives as well as top AI models. The company’s chief scientist and its 30th employee, Jeff Dean, announced his departure after 27 years. He planned to co-found an AI startup called Discovery Loop along with colleagues Sanjay Ghemawat, Quoc Le, and Oriol Vinyals.
Meanwhile, Demis Hassabis stepped down as the CEO of Google DeepMind. He instead became the unit’s chairman, as well as Alphabet’s chief scientist. He will continue to run Isomorphic Labs, while Koray Kavukcuoglu will be the senior vice president tasked with Gemini model development.
Google is facing immense pressure to ship a model that can definitively keep the U.S. ahead of China in the AI race, but the progress of its anticipated Gemini 3.5 Pro model is still unknown.
The search engine giant had said last month that the model was being tested with partners and could come soon.
This week also saw Google launching a new line-up of Pixel phones, packed with the company’s latest Gemini AI integrations. The devices’ upgraded hardware did not offer many radical innovations.
On August 13, Google announced the launch of Gemini 3.7 Flash, which the company described as its most intelligent workhorse model for coding and agents, so far.
“Gemini 3.7 Flash delivers a noticeably improved developer experience over 3.6 Flash. It better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity. It thinks more diligently, putting in more effort into multi-step planning and tool calls,” noted the company.
Anthropic
The Claude-maker certainly made headlines this week. This was not thanks to a new model launch, but due to an announcement about AI labelling. Anthropic announced that in compliance with the European Union’s AI Act, Claude-generated text would carry embedded watermarks, while generated files are set to include digitally signed provenance metadata.
“As AI-generated content becomes commonplace, greater transparency and signals about where content comes from can give people useful context about the information they consume. To support transparency and comply with our legal obligations, Anthropic is working to include machine-readable marks in content that Claude generates,” stated the company in a support page.
However, the AI company backed by Google and Amazon did not provide the technical details of how the watermarking process would apply to text.
Users also raised concerns about the feature that allows Claude to mark text even if the human-made file was only summarised, proofread, converted, or lightly edited with the AI tool.
False positives in AI detection tools can lead to reputational harm and ramifications for the users’ career, while false negatives can enable plagiarism and cheating.
OpenAI
The ChatGPT-maker introduced expansions to Daybreak—a system for accessing models designed for cybersecurity use-cases—as well as a new model reserved for a limited set of users.
“We’re expanding our cybersecurity initiative Daybreak and introducing GPT-5.6-Cyber, a new model for advanced, authorized cybersecurity work. As the threat landscape evolves, we’re putting frontier intelligence in the hands of trusted defenders before attackers can deploy offensive AI at scale,” said OpenAI.
OpenAI reported it had used GPT-5.6-Cyber extensively in real-world vulnerability research, and highlighted that access was limited to approved defenders. There are additional controls and monitoring for higher-risk cybersecurity work, according to the company.
Published – August 14, 2026 06:24 pm IST
