Qwen Gains Ground, Codex Grows: Today’s AI News
Alibaba, OpenAI, and Google deliver new AI models, while China, companies, and web search feel the headwinds. Plus: a fresh agent approach.
Inhaltsverzeichnis
Today it becomes pretty clear where the AI market is headed: more efficient open-source models, increasingly powerful developer tools, and a growing power struggle over data, distribution, and control. At the same time, new research shows that the path to truly usable AI agents is still more of a marathon than a sprint.
🔥 Alibaba launches Qwen3.6-35B-A3B
With Qwen3.6-35B-A3B, Alibaba is releasing a remarkably efficient open-source model: out of 35 billion parameters, only around 3 billion are active at a time. That is not just a nice architectural trick for the slide deck, but above all relevant for costs, latency, and operation on real infrastructure.
What is especially interesting is the benchmark comparison: according to reports, the model outperforms Google’s significantly larger Gemma 4-31B in programming and reasoning. Once again, this shows that “bigger” does not automatically mean “better.” For teams looking for a strong coding or reasoning model without immediately burning through a GPU farm, this is a very good signal. For the rest of the market, it means that efficiency remains a killer feature in the open-source camp.
🧭 Beijing classifies Meta’s Manus deal as a threat
Meta’s planned Manus deal is creating geopolitical headwinds. According to the Financial Times, China’s National Security Commission reportedly sees the acquisition as an attempt to undermine China’s technological base. At the same time, the startup’s founders are said to be being held in the country.
This is more than just another Silicon Valley M&A thriller. The case shows how strongly AI startups are now coming into the focus of security policy. Anyone today with a relevant model, a useful agent product, or access to valuable data is no longer just an acquisition target, but potentially a strategic asset. For founders, that means exit strategies are becoming more complex when states are sitting at the table — or playing along under the table.
🤖 OpenAI upgrades Codex to challenge Claude Code
OpenAI is giving Codex a major upgrade: the tool can now operate the Mac on its own, generate images, remember preferences, and keep working on tasks over longer periods of time. This positions OpenAI even more clearly against Anthropic’s Claude Code.
For developers, this matters because a pattern is emerging here: AI tools are becoming less like “helpful assistants” and more like semi-autonomous work environments. That saves time, but also shifts responsibility. If a tool works on a project for weeks, then good task management becomes almost as important as good prompting. So the real question is not only what Codex can do, but how much control you give up when you let it run. Practical, yes. A little unsettling, too.
🏢 Companies are using AI, but governance is lagging behind
A new analysis from heise shows a familiar pattern: German companies are increasingly using AI, but governance, control, and oversight are still falling short. In short: the technology is arriving, the rules are trailing behind.
This is crucial for day-to-day business operations. Once AI starts affecting processes, customer communication, or internal knowledge work, pilot projects and innovation slides are no longer enough. Clear responsibilities, documented risks, approval workflows, and at least a minimum of auditability are needed. Otherwise, “We use AI” quickly becomes “We hope it’ll be fine.” And as we know, that is not a reliable compliance plan.
🧩 New research: Lighter GUI agents through multi-role orchestration
With Towards Scalable Lightweight GUI Agents via Multi-role Orchestration, a new arXiv paper presents an approach intended to make autonomous GUI agents lighter and more efficient. The idea: instead of one large agent doing everything, a multi-stage architecture takes on different roles and tasks.
This matters because GUI agents have been touted for months as the next major automation step, but in practice they often fail because of latency, energy consumption, and stability. Especially on endpoint devices, “works sometimes” is not a sufficient product category. The new approach targets exactly this hurdle. If such architectures prove themselves, we could see AI assistants that not only talk intelligently, but also run affordably and robustly on normal devices.
🔎 Google makes the classic website visit even less important
Google is integrating its AI Mode more deeply into Chrome. In the future, websites are supposed to appear directly next to the AI answer instead of users first navigating to the site as their primary destination. That sounds like a small UI detail, but strategically it is quite significant.
Because this shifts Google traffic even further away from the classic click to the actual page. For publishers, SEO teams, and media companies, this is the next round in the long game for visibility in the AI age. If answers are created directly in the browser workflow, the content of the website is still needed, but the visit itself continues to lose value. Welcome to the age of “read without being visited.”
🦾 DeepMind launches Gemini Robotics-ER 1.6
Google DeepMind has unveiled an update to its robotics model with Gemini Robotics-ER 1.6. The system is designed to help robots plan and act more precisely and adds, among other things, a new ability to read measurement instruments.
That sounds like a niche feature, but in robotics it is highly relevant. Once machines are deployed in labs, manufacturing, or service environments, it is not only language understanding that matters, but also the ability to translate visual data into concrete actions. This is exactly where the line between a “multimodal model” and a practically usable robot brain begins to blur. It is not a sci-fi butler yet, but it is a pretty solid step toward real physical automation.
🛠️ Tool tip of the day
If you are experimenting today with AI development, agents, or coding workflows, it is worth taking a look at a tool for structured model and prompt workflows: #. Especially with multiple models, tests, and iterative agent setups, it can save you a lot of time — and a few gray hairs.
Don’t want to miss any news? Subscribe to the newsletter