Google Under Pressure, New Models, and AI on Smartphones
EU regulation is hitting Google, while open-weight models, on-device AI, and new tools are visibly reshaping the market today.
Inhaltsverzeichnis
Today’s AI news is mainly about shifts in power: regulation is forcing major platforms to become more open, while new models and tools show how quickly the market is evolving on the technical side. What’s especially exciting: AI is not just getting bigger anymore, but also smaller, more local, and easier to integrate into everyday life.
For you, that means less abstract model theater and more concrete impact on search, Android, on-device AI, and the question of who actually controls the AI interface to the user.
🏛️ EU puts Google under pressure in search and Android
The most interesting regulatory story of the day is the increasing opening of Google’s ecosystem under EU pressure. According to the assessment around the DMA, Google must make search data and Android more accessible to AI competitors — exactly where control over access to the web and to smartphones is especially valuable. This affects not only classic search, but also the question of which assistant even gets a fair chance on Android. Source
Why does this matter? Because AI assistants without data, distribution, and default placement are not magic — they are just software with poor starting conditions. If the EU stays consistent here, this could mean more competition in search, assistants, and mobile in the medium term. In short: less “Gemini everywhere,” more choice — at least on paper.
🤖 Mira Murati’s Inkling: open model with a fine-tuning focus
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, is entering the market for open AI models with Inkling. The model is multimodal, large, and designed for adaptability — less about the flashy “we’re number one” moment and more about practical usefulness in real workflows. That’s exactly what makes it interesting for teams building their own products, agents, or specialized assistants. Source
That makes the open-weights model market even more crowded. For developers, that’s good news because choice increases and pricing pressure rises. For frontier models, it’s uncomfortable because raw benchmark power is no longer the only thing that matters — tooling, adaptability, and price-performance matter too. And yes: when former OpenAI top talent is now pushing open models, that’s at least a small ironic footnote.
📱 Bonsai: 27B model runs directly on the iPhone
With Bonsai 27B, PrismML shows how far on-device AI has come. The 27-billion-parameter model was compressed enough to fit under 4 GB and run directly on an iPhone. According to the company’s own benchmarks, it retains around 90 percent of the original performance, with almost no noticeable loss in math and coding tasks. Source
This matters because on-device AI eases two old problems: latency and privacy. If models run locally, not everything has to be sent to the cloud — a real advantage for mobile assistants, offline scenarios, and sensitive data. This is especially relevant for Apple, since the company has not exactly been known as a sprint champion in local AI so far. Bonsai could show how much more is already possible on smartphones when model compression is taken seriously.
🔐 xAI releases Grok build as open source after data leak
xAI has released its command-line tool “grok” as open source on GitHub after a massive data leak. The trigger doesn’t exactly sound like best practice: the tool apparently uploaded entire directories, including SSH keys and password databases, to cloud servers without asking. That was followed by the usual combination of crisis communication, deletion announcements, and source code publication under Apache 2.0. Source
Why is this more than just an embarrassing incident? Because security culture in AI tools is not a side issue. Anyone building agents or CLI tools that work with files, credentials, and infrastructure needs robust safeguards — otherwise automation quickly becomes a data leak with a speed boost. Open source helps with transparency, but it does not replace solid security. A lesson Silicon Valley apparently prefers to learn live.
🧠 Kimi K3: large open-weight model with lots of context
With K3, Kimi is presenting a new open-weight model that comes with 2.8 trillion parameters and a one-million-token context window. In its own benchmarks, it comes very close to strong Western models and clearly outperforms some competitors. The full weights are expected in the next few days, but independent comparisons are still pending. Source
The exciting news is not just the size, but the signal it sends: open weight from China is no longer a low-cost niche. If these results hold up, the pressure on US and European providers will keep increasing — on price, performance, and accessibility. For ambitious teams, that’s good news, because more strong models mean more options for fine-tuning, agents, and production-ready AI stacks. For providers’ pricing tables, less so.
🔗 Google expands Gemini Notebook into an AI ecosystem
Google is renaming NotebookLM to Gemini Notebook and integrating the tool more deeply into its own ecosystem. Most notably, each notebook gets a cloud computer that can even write and run code — for now, though, only for AI Ultra and Workspace customers. At the same time, apps are being integrated into Google’s AI search. Source
The strategy is pretty clear: Google doesn’t just want to provide a model, but control the entire working environment. Notebook, search, apps, execution — all as much as possible in the same orbit. For users, that can be convenient because research, analysis, and execution move closer together. For competition, it’s another sign that AI platforms are increasingly becoming closed product worlds in which not only models, but also distribution, decide the outcome.
🛠️ Tool tip of the day: Notion AI for structured knowledge work
If you’re juggling multiple sources, notes, and project ideas, a tool like Notion AI can help bundle content and quickly turn chaotic information into usable drafts. Especially when combined with your own docs, meeting notes, and a clear knowledge structure, it’s quite useful for research and content work. #
Want to make sure you don’t miss any news? Subscribe to the newsletter