Alibaba has released Qwen-Image-2.0, a new generation image generation and editing model that unifies 'image generation + image editing' within a lightweight architecture for the first time. It achieves a score of 1029 on the AI Arena text-to-image benchmark and 1034 on image editing tasks. The model supports ultra-long input instructions up to 1K tokens and outputs at up to 2K resolution, emphasizing precise rendering of Chinese characters and multiple fonts. It can generate layout-based assets such as PPTs, infographics, and multi-panel comics, and supports uploading multiple images for editing features like group photos and grid layouts. The API is now available for invitation-only testing via Alibaba Cloud's Bailian platform, while users can also experience it for free through Qwen Chat.

Alibaba DAMO Academy has open-sourced RynnBrain, an embodied foundation model designed for robotic 'mobile manipulation' tasks. It introduces text-spatial interleaved reasoning and spatiotemporal memory mechanisms, using coordinate tokens to ground language decisions onto object positions, reducing hallucination and drift in physical environments. Its MoE version, RynnBrain-30B-A3B, activates approximately 3B parameters during inference. According to materials, it outperforms 72B-class models on several embodied benchmarks and includes the release of RynnBrain-Bench, a fine-grained evaluation benchmark. Model weights, code, and benchmarks are freely available to developers for post-training and engineering deployment in planning, navigation, and related applications.

The Chilean National Center for Artificial Intelligence (CENIA) has led the launch of Latam-GPT, an open-source large language model developed jointly by over 30 institutions from eight Latin American countries. The goal is to correct linguistic and cultural biases arising from English-dominated training data. Trained on more than 8TB of regional private and synthetic data, the model currently supports Spanish and Portuguese, with plans to include indigenous languages in future versions. The project cost approximately $550,000, initially developed using Amazon Web Services, with the next phase to be trained on a supercomputer at Universidad de Tarapacá in northern Chile. The team states that the model and associated resources are openly accessible to support localized applications and regulatory research.

AI video generation company Runway has announced the completion of a $315 million Series E funding round led by General Atlantic, with participation from Nvidia, Adobe Ventures, AMD Ventures, and others. The post-money valuation is approximately $5.3 billion, up from around $3.3 billion in the previous round. The company stated that funds will be used to pre-train the next-generation 'world models,' expand enterprise products and compute infrastructure, and further develop Gen-4.5's capabilities in multi-shot consistency, character persistence, and native audio integration. Runway also disclosed collaboration with CoreWeave to scale computing capacity and plans to grow its team and R&D investment. Customers reportedly span film, advertising, gaming, and enterprise content production.

Google Cloud has launched Gemini 3 Flash, optimized for high-frequency enterprise workflows with emphasis on low latency and cost efficiency. It is now available on Gemini Enterprise and Vertex AI, supporting multimodal understanding, code agent operations, and real-time responsive applications. Case studies cited include Box reporting a 15% improvement in overall accuracy, Harvey achieving over 7% gains in legal tasks, and Geotab showing about 10% improvement on agent coding tasks. The model addresses use cases such as real-time video analysis, structured document extraction, intelligent customer service, and automated prototyping—targeting enterprise applications requiring scalable usage and cost control.

Materials indicate that OpenAI began testing ad placements on ChatGPT’s free and Go subscription tiers in the U.S. starting February 11. Ads will be clearly labeled, and OpenAI states they will not affect model responses. Users may opt to remove ads via paid subscriptions or choose to reduce their free message quota in exchange for an ad-free experience. This test marks a strategic expansion of ChatGPT’s monetization from a pure subscription model toward a 'subscription + advertising' approach. It raises upcoming discussions on privacy boundaries in ad targeting, rules governing the use of conversation content, and potential commercial and compliance implications involving content ecosystems such as media and publishers.

GitHub has previewed support for Claude and OpenAI Codex coding agents, available to Copilot Pro+ and Enterprise users. Multiple agents can run in parallel across GitHub Web, GitHub Mobile, and Visual Studio Code to perform coding, bug fixing, and refactoring tasks. The update coincides with growing concerns over AI-generated code and auto-commits, prompting GitHub to explore enhanced pull request governance and review mechanisms. More secure execution environments, such as sandboxes, are being introduced to mitigate security and supply chain risks from generated code. This integration embeds 'agent-driven development' deeper into primary developer workflows, introducing new engineering requirements for team collaboration and permission management.

Security firm Praetorian has released Augustus, an open-source tool aimed at advancing large language model (LLM) security testing from research demonstrations to repeatable production practices. The tool supports both cloud-hosted and locally deployed models and enables parallel execution to increase testing throughput. It covers various adversarial probing techniques, including prompt rewriting, translation, and encoding transformations, simulating attacker 'variant' strategies to bypass defenses. Augustus can be integrated into CI/CD pipelines and red team workflows, helping enterprises and managed service providers (MSPs) continuously validate model and application defenses, track behavioral changes, and support compliance due diligence, thereby reducing security regression risks during version updates.

Security firm PromptArmor warns that when AI agents integrate with messaging apps supporting 'link previews' (such as Slack, Telegram, etc.), zero-click prompt injection attacks could lead to data leaks. Attackers may trick agents into generating malicious URLs containing sensitive information; automatic preview fetching by the app triggers requests that send keys and other data to attacker-controlled servers—without user interaction. Microsoft Teams combined with Copilot Studio is reported to have the broadest exposure, with additional risks noted in Discord and OpenClaw configurations. Recommendations include providing finer-grained preview controls in messaging apps and advising developers to avoid such integrations in high-security environments while auditing agents’ external communication capabilities.

The Irish government has published the general scheme of the 'Artificial Intelligence Regulation Act 2026' to implement the EU AI Act. The framework establishes distributed enforcement: 13 sectoral regulators—including the Data Protection Commission and Central Bank—will oversee AI systems within their domains, while a statutory Irish AI Office will serve as the coordination center, required to become operational by August 1, 2026. Regulators may conduct on-site inspections, request supply chain information, and, as a last resort, demand access to source code of high-risk AI systems. Serious violations may incur fines up to €35 million or 7% of global annual revenue, whichever is higher. Enforcement decisions require independent adjudication and confirmation by the High Court before taking effect.