AI News Roundup: Liquid AI, Gemini 4 Argon, and Anthropic Warnings
Daily AI news: Liquid AI releases on-device model, Gemini 4 Argon leads robotics spatial reasoning, Anthropic warns of smart home worms, and Kimi K3 raises prices.
👁 112 reads
Today in AI brought specialized new models, critical security warnings, and major investments in sovereign tech infrastructure. Liquid AI and Google DeepMind introduced focused architectures, Anthropic sounded the alarm on an emerging class of threats, and France rolled out coordinated deployments across multiple operational sectors.
Liquid AI Focuses on Local Agents with New Hugging Face Release
Liquid AI has released LFM2.5-24B-A2B on Hugging Face, a 24-billion parameter model featuring an active parameter space of 2 billion. Engineered specifically for fast, private agentic workloads running directly on user hardware, it serves as a lightweight alternative for organizations that want to deploy AI agents without routing sensitive data through external cloud services.
Google DeepMind Expands Robotics Push with Gemini 4 Argon
Google DeepMind has introduced Gemini 4 Argon, a specialized model tailored for robotic applications. Testing on the ERQA spatial reasoning benchmark saw the model achieve a score of 87%, outpacing all rival frontier systems. The release highlights DeepMind’s deliberate strategy to secure leadership in physical and spatial intelligence.
Anthropic Warns AI Worms Could Soon Target Smart Home Devices
Researchers at Anthropic caution that small AI models could begin autonomously self-replicating across internet-connected hardware within months, behaving similarly to traditional computer worms. Connected smart home hardware is identified as the primary attack surface, marking one of the most pressing infrastructure safety concerns raised by frontier labs to date.
Pentagon Replaces Anthropic AI Without Operational Disruption
Palantir’s chief technology officer confirmed on CBS television that the Pentagon has replaced Anthropic models with alternative systems. According to the interview, the transition took place without disrupting ongoing operations, representing an unusually smooth swap of core AI components within a high-stakes defense setting.
Defunct Startups Become New Target for AI Training Datasets
AI labs have started acquiring archival assets from bankrupt startups-including corporate emails, internal chat logs, and proprietary codebases-as conventional public and synthetic datasets hit practical utility limits. Digital remnants of failed ventures are emerging as a lucrative data market, prompting fresh questions around corporate privacy and post-insolvency data governance.
Kimi K3 Raises API Pricing with Major Increase to Output Tokens
The provider behind Kimi K3 has adjusted its pricing structure. Output tokens now cost $13.50 per million, up from $10.00-a 35 percent increase on the most cost-sensitive line item in production environments. Input pricing remains set at $0.73 per million tokens.
Read more
France Launches Sovereign AI Deployments from Orbit to Highways
The French AI ecosystem has advanced from planning to field deployment, with NVIDIA announcing joint projects spanning local server production, an in-orbit autonomous satellite, and automated agents managing nearly 6,000 kilometers of departmental roads. The rollout forms one of the most comprehensive instances of sovereign AI infrastructure development in Europe.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →More Updates
That wraps up the primary developments from the past 24 hours. For complete ongoing industry coverage, visit our latest updates section.
