Created automatically with AI: the texts and images in this edition are AI-generated, based on the linked original sources. Errors may occur.
Mistral and Google unveil new models for language processing, search and image generation. Meanwhile, there is news about AI safety, autonomous agents, and their use in defense and medicine.
AI-generated image
01heise online
Mistral unveils Large 4 and plans to release open weights
Mistral has unveiled Large 4 as a new language model. According to the report, it has a total of one trillion parameters, of which 49 billion are active. It performs competitively on selected tests of coding and agentic tasks, according to comparisons published by the maker; however, leading models from OpenAI and Anthropic are well ahead in software development. The weights are to be released at the end of October, following a safety review.
Why it matters
The model expands the range of powerful language models from a European provider. Some of the published comparisons are based on claims by the maker and do not allow for a comprehensive performance comparison with all leading models.
Google releases multimodal embedding model for local search
Google has unveiled EmbeddingGemma 2, with 740 million parameters. The open model maps text, images, audio, video and code into a shared vector space so that content can be searched and compared semantically. According to Google, it is designed to run locally and can be combined with generative models such as Gemma 4 for on-device retrieval-augmented generation applications. The weights are available under the Apache 2.0 license.
Why it matters
Local search and data processing can enable applications that find content across different media without sending it to external servers. Performance claims and benefits of use come from the provider.
Google launches cheaper image generation with Nano Banana 2.1
Google has released Nano Banana 2.1 for image generation and editing. According to the report, the model is based on Gemini 3.6 Flash and is offered at lower prices than its predecessor. Google cites improvements in image quality, text rendering and character consistency, among other things. However, a comparison of published tests does not consistently favor the new version; Nano Banana Pro often produces better images in practice.
Why it matters
The model expands Google’s image tools and, according to the stated prices, lowers the cost per generated image. Quality claims are based on statements from the maker and tests whose results do not present a consistent picture.
Musubi has unveiled PolicyLM-1.7B, with open weights, for content moderation. The model is intended to apply rules written in natural language to posts and, according to the company, deliver a binary judgment in less than 50 milliseconds. Musubi says policies can be changed without retraining the model. The approach is intended to combine the speed of conventional moderation classifiers with the flexibility of modern language models; the report contains no independent performance data.
Why it matters
An adaptable model could help platforms change moderation rules without having to train a new model for each change. Claims about speed and usefulness are, so far, those of the provider.
Anthropic tiers access to powerful cybersecurity capabilities
Anthropic is expanding its Cyber Verification Program and dividing access into three tiers. Vetted organizations can receive capabilities for defensive security work, authorized red-team testing or limited specialized tests, depending on the use case. Higher-tier access requires additional reviews and controls; certain dangerous actions are still to be blocked. Data retention is generally planned. Anthropic does, however, cite an exception for organizations with zero-data-retention access to Claude Fable 5.1 or Claude Mythos 5.1.
Why it matters
The program is intended to give security teams access to capabilities that are more restricted in generally available models. Anthropic itself describes the access requirements and safeguards, so their effectiveness has not been independently established.
European Parliament approves €115 million program for defense innovation
The European Parliament has approved the Agile pilot program, with a budget of €115 million. It is intended to support start-ups and medium-sized companies working on defense technologies, including AI-powered reconnaissance and robotics. Lethal autonomous weapons systems without effective human control are excluded from funding. The program still requires formal approval by the Council of Ministers; it is scheduled to launch on January 1, 2027.
Why it matters
Agile establishes a funding framework for military applications of AI and other technologies, while also drawing a line at autonomous lethal weapons systems. The initiative is not yet definitively in force.
Tech companies work on standard for AI agents’ access to websites
Meta, Walmart, Stripe and other companies are working, according to TechCrunch, on an open standard for communication between AI agents and businesses. In online retail, it is intended to help distinguish authorized agents from unwanted automated access. Agents currently sometimes run into deliberately deployed blocks or conventional anti-bot checks; such barriers can also be triggered unintentionally. The report provides no details on implementation or a timeline.
Why it matters
A shared standard could affect whether and how personal AI agents can make purchases or bookings on websites. It also concerns the protection of platforms against unauthorized automated access.
Utah approves limited pilot for AI prescriptions for acne
The US state of Utah has granted start-up Nolla Health an exemption to run a pilot. The app may recommend and prescribe certain topical medications for adults with mild to moderate acne; more severe cases, pregnant women and oral medications are excluded. In the first phase, two doctors must approve each prescription in advance. Later, post-prescription checks and random audits are planned. The approval is initially valid for 12 months.
Why it matters
The trial tests how far an AI may act independently in a narrowly defined medical field. The staged physician reviews and reporting requirements are part of the regulatory conditions.
Automatically summarized with AI and checked against the source texts retrieved. Original sources are linked for every story; errors may still occur. An editorial overview from Medienfuchs.
