AI NEWS
From verified sources. Refreshed every 2 hours.
Interfaze is a hybrid AI model that combines task-specific neural networks (for OCR, object detection, speech recognition) with a transformer decoder through a shared embedding space. The architecture achieves competitive or superior performance on specialized benchmarks compared to larger generalist models while operating at lower computational cost by activating only relevant parameters per task.
Read on arXiv cs.AI →SAM 3D is a generative model that reconstructs 3D objects from single images by predicting geometry, texture, and layout. The researchers created a large-scale annotated dataset through human-model collaboration and trained the system using synthetic pretraining with real-world alignment, achieving substantial improvements over existing methods on natural, cluttered scenes.
Read on arXiv cs.AI →NVIDIA released Nemotron 3.5 Content Safety, a multimodal safety system designed to help enterprises customize content filtering for AI applications across different regions and use cases. The tool enables organizations to establish guardrails aligned with their specific policies and cultural requirements.
Read on Hugging Face →NVIDIA's Nemotron 3 Ultra, a 550-billion-parameter open language model, is now available on Amazon SageMaker JumpStart with one-click deployment. The model uses a hybrid Transformer-Mamba architecture and claims 5x faster inference and 30% lower costs for autonomous agent workloads.
Read on AWS Machine Learning →OpenAI has implemented a memory system for ChatGPT that allows the model to retain user preferences and maintain context across multiple conversations, potentially improving the relevance and personalization of responses over time.
Read on OpenAI →Fundamental's NEXUS, a foundation model designed specifically for tabular data prediction, is now available on Amazon SageMaker JumpStart. The model leverages pre-training on billions of structured datasets to enable enterprises to generate accurate predictions from their data without extensive feature engineering, reducing deployment timelines from months to days.
Read on AWS Machine Learning →Google's new Gemini AI agent Spark demonstrates advanced capabilities by accessing personal information about users without explicit disclosure, raising concerns about privacy and the broader focus on productivity gains rather than addressing substantive societal challenges.
Read on The Verge AI →OpenAI has introduced GPT-Rosalind, a specialized AI model designed to support life sciences research through improved biological reasoning, drug chemistry analysis, genetic data interpretation, and automation of lab procedures.
Read on OpenAI →Nvidia has introduced RTX Spark chips designed for laptops that could make practical AI-capable personal computers viable. The chips aim to address previous limitations in bringing AI capabilities to consumer laptops.
Read on Wired AI →NVIDIA introduced Cosmos 3, a multimodal world model that processes and generates language, images, video, audio, and actions in a single unified framework. The system achieved top performance on text-to-image, image-to-video, and robotic control benchmarks, with code and models released openly to support physical AI research.
Read on arXiv cs.AI →Researchers introduced TypewriterLM, a 7.24B parameter language model trained exclusively on pre-1913 English text, along with TypewriterCorpus, a 54B-token historical dataset. The work addresses challenges specific to historical language modeling through specialized data cleaning, leakage prevention, and evaluation benchmarks designed to assess temporal consistency.
Read on arXiv cs.AI →Microsoft announced MAI-Thinking-1, a new advanced reasoning model developed in-house, at Build 2026. The medium-sized model performs comparably to leading models on software engineering benchmarks and was trained on clean data without relying on third-party model distillation, marking Microsoft's continued expansion of its own AI model capabilities.
Read on The Verge AI →Microsoft has released Scout, a new AI assistant designed to integrate OpenClaw-inspired capabilities into Microsoft 365. The assistant was unveiled at Microsoft Build and aims to provide users with enhanced power and flexibility within the Microsoft 365 ecosystem.
Read on TechCrunch AI →Microsoft launched Scout, a new AI personal assistant built on OpenClaw that integrates with Microsoft 365 applications including Outlook, Teams, and OneDrive. Unlike existing Copilot features, Scout functions as a dedicated assistant capable of handling calendar management, expense reporting, email composition, and other workplace tasks.
Read on The Verge AI →Microsoft unveiled Project Solara, a new operating system designed for AI agent-powered devices, built on Android rather than Windows. The company demonstrated two concept devices at Build 2026: a desk-mounted display with facial recognition and a wearable badge with camera and fingerprint scanner.
Read on The Verge AI →OpenAI released new capabilities for its Codex tool designed to expand its applications in enterprise workplace settings. The company also published internal research documenting how Codex is being utilized for knowledge work tasks.
Read on TechCrunch AI →Hugging Face released Holo3.1, a new computer use agent designed for fast, local operation. The model enables autonomous task completion on computers while running on-device without requiring external services.
Read on Hugging Face →Google launched Gemini Spark, an agentic AI tool designed to handle multi-step tasks like trip planning. The reviewer found it delivered more sophisticated results than previous AI assistants, moving beyond generic recommendations to provide more personalized and detailed planning capabilities.
Read on The Verge AI →Researchers developed Ryze, an automated system that converts biomedical papers into evidence-enriched training data for specialized vision-language models. The resulting BioVLM-8B model significantly outperforms general-purpose models on biomedical benchmarks by preserving evidence connections across document elements, and both the system and model are being released open source.
Read on arXiv cs.AI →OpenAI's GPT-5.5, GPT-5.4, and Codex models are now generally available on Amazon Bedrock for production use. Pricing for GPT-5.5 matches OpenAI's direct rates, while Codex uses pay-per-token pricing integrated with AWS billing commitments.
Read on AWS Machine Learning →Nvidia is entering the consumer laptop chip market with RTX Spark, positioning itself to deliver the performance breakthrough that Windows laptops have lacked compared to Apple's Arm-based chips. The move could replicate Apple's M1 success in the Windows ecosystem, though the technology is expected to come at a premium price point.
Read on The Verge AI →Google launched Gemini Spark, an AI agent that performs multi-step tasks autonomously in the background with user oversight. The reviewer found it capable but questioned whether the cost and privacy implications justify adoption.
Read on The Verge AI →Windborne Systems has developed a weather forecasting model that outperforms government agencies' predictions, achieving superior accuracy over multi-day timeframes. The advancement could reshape weather prediction capabilities in both private and public sectors.
Read on TechCrunch AI →JetBrains has released Mellum2, a 12-billion parameter mixture-of-experts model. This lightweight MoE architecture aims to provide efficient performance for various AI tasks.
Read on Hugging Face →NVIDIA has released Cosmos 3, an open-source foundational model designed for physical AI applications that combines reasoning and action capabilities. The model aims to enable systems to understand and interact with physical environments, marking a step toward practical AI agents.
Read on Hugging Face →Researchers introduced AMix-2, a foundation model that treats protein sequences as a native modality alongside text in large language models, enabling unified protein understanding and design tasks. The model uses block-wise diffusion language modeling and was evaluated on ProteinArena, a new benchmark with realistic generalization protocols, where it outperformed existing LLMs and matched specialized protein models.
Read on arXiv cs.AI →Google published 11 demonstration videos showcasing the capabilities of Gemini Omni and Gemini 3.5, two AI models announced at Google I/O 2026. The demos illustrate practical applications and features of these newly released models.
Read on Google AI Blog →Researchers introduced AgentDoG 1.5, a lightweight safety alignment framework for AI agents that addresses emerging risks from autonomous systems. The framework includes compact models (0.8B-8B parameters) trained on minimal data, an efficient deployment system reducing computational overhead, and real-time safety moderation capabilities for agent operations.
Read on arXiv cs.AI →Researchers introduced Aryabhata 2, a language model fine-tuned with reinforcement learning to solve competitive STEM exam problems from Indian entrance exams like JEE and NEET. The model outperforms its base version while using significantly fewer tokens, addressing the need for scalable, reliable STEM problem-solving systems.
Read on arXiv cs.AI →OpenAI has expanded access to GPT-Rosalind, a specialized AI model designed for biodefense applications, making it available to approved developers and U.S. government agencies working on public health and pandemic preparedness initiatives.
Read on OpenAI →Anthropic's Claude Opus 4.8 model is now available through Amazon Bedrock and Claude Platform on AWS. The update offers improved performance for coding, autonomous tasks, and long-running workflows while maintaining enterprise security and data residency options.
Read on AWS Machine Learning →Anthropic launched an updated Opus model featuring Dynamic Workflows, a new capability that enables coordination of multiple AI subagents working together. This tool allows for more complex task execution through orchestrated agent collaboration.
Read on TechCrunch AI →Anthropic is releasing Claude Opus 4.8, designed to better acknowledge uncertainty and avoid making unsupported claims. The model shows approximately 4x less tendency than its predecessor to present conclusions confidently without sufficient evidence.
Read on The Verge AI →Researchers developed Soro, a lightweight language model optimized for Tajik language use cases, building on Gemma 3 and trained on 1.9 billion Tajik tokens. The model includes new Tajik benchmarks for evaluation and supports quantized versions for deployment in resource-constrained environments, with pilots planned for Tajikistan's education sector.
Read on arXiv cs.AI →Researchers introduced Laguna M.1 and Laguna XS.2, two mixture-of-experts models designed for coding tasks with 225.8B and 33.4B parameters respectively. Both models were developed using an integrated system called Model Factory and demonstrate competitive performance on software engineering benchmarks, with the smaller XS.2 model released as open source.
Read on arXiv cs.AI →Researchers introduced BlazeEdit, a lightweight image editing diffusion model with only 195M parameters designed for mobile devices. By removing text-conditioning and consolidating multiple editing tasks into one model, it performs inference in 290ms on smartphones while preserving privacy and maintaining competitive quality.
Read on arXiv cs.AI →Researchers developed SafeMed-R1, a medical language model trained with clinician-reviewed reasoning chains and safety alignment techniques. The model achieved 79.6% accuracy on clinical benchmarks and demonstrated reduced unsafe outputs under adversarial testing, matching junior resident performance on medication safety tasks.
Read on arXiv cs.AI →ElevenLabs released a music generation model that allows users to regenerate specific sections of a song while preserving the rest of the track. This enables mid-track genre switching and targeted edits without re-creating entire compositions.
Read on TechCrunch AI →MiniMax released the M2 series of Mixture-of-Experts language models with 229.9B total parameters but only 9.8B activated per token. The models are optimized for agentic tasks using agent-driven data pipelines and a specialized reinforcement learning system, with the latest M2.7 checkpoint showing capabilities for autonomous self-improvement.
Read on arXiv cs.AI →Researchers introduced JT-Safe-V2, a language model emphasizing safety through design improvements including contextual pre-training data and enhanced post-training mechanisms. The team also proposed Safe-MoMA, a framework using multiple models to reduce inference costs by over 30% while maintaining performance, and released the 35B parameter model publicly.
Read on arXiv cs.AI →Google's Gemini model demonstrates advanced video generation capabilities that can create realistic content with minimal effort, raising questions about the accessibility of AI-generated media and potential misuse concerns.
Read on The Verge AI →Amazon Nova Act, AWS's browser-based AI agent service, now qualifies as HIPAA-eligible, enabling healthcare organizations to automate compliance-sensitive workflows involving protected health information. This certification removes regulatory barriers for deploying autonomous agents in healthcare settings for tasks like claims processing and referral coordination.
Read on AWS Machine Learning →Allen Institute released an updated version of OlmoEarth, a family of computer vision models designed for analyzing satellite and aerial imagery. The v1.1 update focuses on improving efficiency while maintaining performance for Earth observation tasks.
Read on Hugging Face →Google announced at I/O 2026 a shift toward agentic capabilities in Gemini, its AI system. The company plans to enable Gemini to perform more autonomous tasks and help users accomplish complex work more efficiently.
Read on Google AI Blog →Google announced Gemini 3.5, its latest AI model series designed to integrate advanced reasoning capabilities with the ability to take actions. The release occurred at Google I/O and represents the company's effort to advance beyond purely analytical AI.
Read on Google AI Blog →Hugging Face introduced the Ettin Reranker Family, a collection of reranking models designed to improve search and retrieval quality by rescoring documents or passages. These models help refine results from initial retrieval systems to provide more relevant outputs.
Read on Hugging Face →Google DeepMind introduced Gemini Omni, a new AI model designed to process and understand multiple types of content including text, audio, and video in an integrated manner. This multimodal capability represents an advancement in the company's AI offerings.
Read on Google DeepMind →Google DeepMind introduced Gemini 3.5, a new AI model designed to execute complex, autonomous workflows. The model emphasizes agent capabilities for handling multi-step tasks.
Read on Google DeepMind →Databricks has integrated GPT-5.5 into its enterprise agent workflows following the model's superior performance on the OfficeQA Pro benchmark. This integration enables businesses to deploy advanced AI capabilities for automation tasks.
Read on OpenAI →IBM released Granite Embedding Multilingual R2, an open-source embedding model under Apache 2.0 license that supports multiple languages and handles 32K token context windows. The model achieves competitive retrieval performance while maintaining under 100 million parameters, making it suitable for efficient multilingual semantic search applications.
Read on Hugging Face →OpenAI has implemented safety improvements to ChatGPT that enhance its ability to understand context within sensitive conversations. The updates aim to better detect potential risks over extended interactions and generate more appropriate responses accordingly.
Read on OpenAI →OpenAI has launched GPT-5.5 and GPT-5.5-Cyber models through its Trusted Access for Cyber program, providing verified security researchers and defenders with advanced AI tools to accelerate vulnerability research and strengthen critical infrastructure protection.
Read on OpenAI →OpenAI has introduced new voice models for its API that support reasoning, translation, and transcription capabilities, enabling more sophisticated voice-based interactions. These models aim to provide more natural and intelligent conversational experiences.
Read on OpenAI →vLLM released a major update transitioning from v0 to v1, emphasizing correctness validation before applying corrections in reinforcement learning workflows. The upgrade focuses on improving the reliability and accuracy of the inference engine's handling of RL-based model refinement.
Read on Hugging Face →Google DeepMind introduced AlphaEvolve, a coding agent powered by Gemini that applies algorithmic solutions to problems across business, infrastructure, and scientific domains. The system demonstrates how AI-driven code generation can address practical challenges in multiple fields.
Read on Google DeepMind →OpenAI released GPT-5.5 Instant as ChatGPT's new default model, featuring improved accuracy, fewer hallucinations, and enhanced personalization capabilities for users.
Read on OpenAI →OpenAI released a system card for GPT-5.5 Instant, providing technical documentation and safety assessment details for this model variant. The document outlines the model's capabilities, limitations, and recommended usage guidelines.
Read on OpenAI →IBM's Granite 4.1 large language models have been released, with documentation detailing their architecture and training methodology. The models represent an update to IBM's open-source LLM line designed for enterprise applications.
Read on Hugging Face →NVIDIA released Nemotron 3 Nano Omni, a multimodal AI model designed to process documents, audio, and video with extended context capabilities. The model enables applications like document analysis agents and multimedia understanding systems.
Read on Hugging Face →OpenAI has made its GPT models, Codex code generation tool, and Managed Agents available through Amazon Web Services, allowing enterprises to deploy these AI systems within their AWS infrastructure for improved security and integration.
Read on OpenAI →DeepSeek released V4, a model variant capable of processing approximately one million tokens in context. The extended context window enables AI agents to work with significantly larger amounts of information in a single interaction.
Read on Hugging Face →OpenAI released GPT-5.5, a new model positioned as their most advanced to date, offering improved performance and speed for specialized applications including software development, scientific research, and analytical work.
Read on OpenAI →OpenAI released a system card for GPT-5.5, a document detailing the model's capabilities, limitations, and safety considerations. The system card provides technical specifications and responsible use guidelines for developers and researchers.
Read on OpenAI →OpenAI released ChatGPT Images 2.0, an upgraded image generation model featuring better text rendering within images, support for multiple languages, and enhanced visual understanding capabilities.
Read on OpenAI →OpenAI launched GPT-Rosalind, a specialized AI model designed to support life sciences research applications including drug discovery, genomic analysis, and protein structure reasoning.
Read on OpenAI →Google DeepMind released Gemini 3.1 Flash TTS, an audio model featuring granular audio tags that enable precise control over speech generation characteristics. This advancement allows developers to create more expressive and customizable AI-generated audio.
Read on Google DeepMind →OpenAI has expanded its Trusted Access for Cyber program by introducing GPT-5.4-Cyber, a specialized AI model made available to approved cybersecurity professionals. The expansion aims to provide verified defenders with advanced AI capabilities while implementing enhanced safety measures as AI cybersecurity tools become more powerful.
Read on OpenAI →Google DeepMind released Gemini Robotics-ER 1.6, an updated version of their robotics model designed to improve how robots understand and interact with physical environments. The upgrade focuses on better spatial reasoning and the ability to process information from multiple camera angles simultaneously.
Read on Google DeepMind →Waypoint-1.5 is a new system designed to create high-quality interactive virtual environments that can run on standard consumer GPUs. The release enables more accessible development of detailed 3D worlds without requiring expensive hardware.
Read on Hugging Face →Google DeepMind released Gemma 4, an open-source model designed for advanced reasoning and agentic workflows. The company claims these models offer improved capabilities compared to previous versions while remaining open for public use.
Read on Google DeepMind →Google released Gemma 4, a new multimodal AI model designed to run on local devices. The model represents an advancement in frontier-class AI capabilities while maintaining efficiency for on-device deployment.
Read on Hugging Face →Falcon Perception appears to be a new model or capability release, though limited details are available. Without the full excerpt, the specific features, applications, or significance cannot be accurately summarized.
Read on Hugging Face →IBM released Granite 4.0 3B Vision, a compact multimodal model designed for enterprise document processing. The lightweight model combines vision and language capabilities while maintaining efficiency for business applications.
Read on Hugging Face →Google DeepMind released an updated voice model with improved accuracy and reduced latency to enhance the naturalness and responsiveness of voice interactions. The upgrade aims to make audio-based AI conversations more reliable and fluid for users.
Read on Google DeepMind →Google DeepMind released Lyria 3 Pro, an upgraded music generation model capable of creating longer tracks with improved structural understanding. The company is expanding Lyria's availability across additional Google products and platforms.
Read on Google DeepMind →OpenAI has designed Sora 2 and its associated app with built-in safety measures to address risks from advanced video generation and social sharing capabilities. The company emphasizes concrete protections as foundational to the platform's architecture.
Read on OpenAI →Holotron-12B is a new computer use agent model with 12 billion parameters designed for high-throughput automated interaction with computer interfaces. The model enables efficient processing of tasks that require understanding and acting on visual desktop environments.
Read on Hugging Face →OpenAI released two smaller variants of GPT-5.4 designed for efficiency and speed. The mini and nano versions target coding tasks, tool integration, multimodal reasoning, and high-volume API applications.
Read on OpenAI →OpenAI has added interactive visual explanations to ChatGPT for mathematics and science topics, enabling students to dynamically explore formulas, variables, and scientific concepts within the platform.
Read on OpenAI →Hugging Face released LeRobot v0.5.0, an update focused on expanding capabilities across multiple dimensions of the robotic learning platform. The release aims to improve scalability and functionality for robotics applications.
Read on Hugging Face →OpenAI released a system card for GPT-5.4 Thinking, documenting the model's capabilities, limitations, and safety considerations. The document provides technical details about the system's reasoning processes and performance benchmarks.
Read on OpenAI →OpenAI released GPT-5.4, a new frontier model designed for professional applications. The model features improved capabilities in coding, computer use, and tool integration, along with support for 1 million token context windows.
Read on OpenAI →Google DeepMind released Gemini 3.1 Flash-Lite, a new model optimized for speed and cost efficiency within the Gemini 3 series. The lightweight variant aims to enable faster inference and reduced computational expenses for deployments at scale.
Read on Google DeepMind →OpenAI released a system card for GPT-5.3 Instant, documenting the model's capabilities, limitations, and safety considerations. The documentation provides transparency regarding the model's performance characteristics and potential risks.
Read on OpenAI →OpenAI released GPT-5.3 Instant, an updated model designed for improved performance in everyday conversational tasks with enhanced fluency and utility. The model aims to provide smoother interactions for typical use cases.
Read on OpenAI →Google DeepMind released Nano Banana 2, an image generation model designed to balance advanced capabilities with fast processing speeds. The model features improved world knowledge, production-ready specifications, and consistent subject rendering while maintaining rapid inference times.
Read on Google DeepMind →Google DeepMind released Gemini 3.1 Pro, an AI model built to handle complex tasks requiring in-depth analysis rather than straightforward answers.
Read on Google DeepMind →Google DeepMind has integrated its Lyria 3 music generation model into the Gemini app, allowing users to create 30-second music tracks by providing text descriptions or images as input.
Read on Google DeepMind →GPT-5.2 generated a previously unknown formula for gluon amplitudes in theoretical physics, which researchers subsequently proved and confirmed. This demonstrates the model's capability to contribute to scientific discovery in advanced physics domains.
Read on OpenAI →Google DeepMind has updated Gemini 3 Deep Think, a specialized reasoning mode designed to tackle complex science, research, and engineering problems. The enhancement aims to improve the model's capability to address contemporary challenges in these technical domains.
Read on Google DeepMind →OpenAI released GPT-5.3-Codex-Spark, a new coding model optimized for real-time performance with 15 times faster code generation and 128k token context window. The model is available as a research preview for ChatGPT Pro subscribers.
Read on OpenAI →Google DeepMind released research demonstrating Gemini Deep Think's effectiveness in accelerating mathematical and scientific discovery across multiple fields. The findings show the model's capability in tackling complex problem-solving tasks within research domains.
Read on Google DeepMind →OpenAI released GPT-5.3-Codex, an agent designed specifically for coding tasks that combines advanced code generation capabilities with broad reasoning abilities for complex technical projects over extended workflows.
Read on OpenAI →OpenAI released GPT-5.3-Codex, a coding model that integrates advanced coding abilities from GPT-5.2-Codex with reasoning and professional knowledge capabilities. The company positions it as the most capable agentic coding model available.
Read on OpenAI →H Company has released Holo2, a new model that demonstrates improved performance in user interface localization tasks. The model appears to represent an advancement in handling multilingual UI adaptation based on its positioning as a market leader in this specialized area.
Read on Hugging Face →Hugging Face released Differential Transformer V2, an updated version of their differential transformer architecture. The advancement likely represents improvements to the model's efficiency or performance characteristics, though specific enhancements are not detailed.
Read on Hugging Face →Overworld released Waypoint-1, a video diffusion model capable of generating real-time interactive video content. The model enables dynamic video creation with interactive elements, advancing capabilities in procedural video generation.
Read on Hugging Face →OpenAI has launched ChatGPT Go worldwide, providing users with broader access to GPT-5.2 Instant, increased usage allowances, and extended context memory at a lower cost point.
Read on OpenAI →Google DeepMind released Veo 3.1, an updated video generation model that produces more natural and engaging video content while adding support for vertical video formats. The update emphasizes improved consistency, creativity, and user control over generated videos.
Read on Google DeepMind →Salesforce launched a rebuilt Slackbot AI agent that transforms the tool from a basic notification system into a fully functional agent capable of searching enterprise data, drafting documents, and executing tasks. The new version, available to Business+ and Enterprise+ customers, represents Salesforce's major push into agentic AI as it competes with Microsoft and Google in workplace AI solutions.
Read on VentureBeat AI →