AI NEWS
From verified sources. Refreshed every 2 hours.
Microsoft released its inaugural AI security model alongside a new platform designed for agentic cybersecurity operations. The releases expand Microsoft's capabilities in applying artificial intelligence to detect and respond to security threats.
Read on TechCrunch AI →Chinese startup Moonshot released Kimi K3, an open-weight AI model that reportedly matches or exceeds performance of leading US models while being significantly cheaper to develop. The free release of model weights to developers has raised concerns in the US tech industry about maintaining competitive advantages through proprietary systems.
Read on The Verge AI →NVIDIA introduced Cosmos-H-Dreams, a generative simulation model designed for surgical robotics that enables real-time video generation. The technology aims to improve training and planning for robotic surgical systems by simulating complex surgical scenarios.
Read on Hugging Face →Anthropic released Claude Opus 5, its most advanced Opus model, now available on Amazon Bedrock and AWS. The model delivers improvements in coding, knowledge work, visual understanding, and long-running tasks while matching higher-tier model performance at Opus pricing, with enterprise features including zero data retention by default.
Read on AWS Machine Learning →Anthropic has released Opus 5, a new AI model that offers improved cost efficiency and fewer operational restrictions compared to its predecessor Fable, potentially making it more attractive for widespread applications.
Read on TechCrunch AI →Anthropic launched Claude Opus 5, a new model that performs comparably to its higher-tier Fable 5 in many areas while showing particular improvements in complex coding. The release comes after regulatory scrutiny of Fable 5 and follows recent security incidents in the AI industry.
Read on The Verge AI →Meta has upgraded its AI chatbot with new productivity capabilities, including calendar integration for event planning, daily briefings, and interactive research features. The company deployed its Muse Spark 1.1 model to enhance the assistant's functionality as it competes with ChatGPT, Gemini, and Claude.
Read on The Verge AI →OpenAI's GPT-5.6 models (Sol, Terra, and Luna) are now available on Amazon Bedrock, offering developers access to frontier models through AWS infrastructure for various workloads including coding agents, reasoning tasks, and high-volume inference with AWS security and cost controls.
Read on AWS Machine Learning →OpenAI has launched voice mode for ChatGPT's desktop application, enabling users to interact with ChatGPT Work and Codex through voice commands. The feature allows voice-based task completion and agent control directly within the desktop environment.
Read on TechCrunch AI →Anthropic has enhanced Claude's voice mode with improved capabilities, enabling users to perform tasks like rescheduling meetings and drafting emails through voice interaction. The update leverages more advanced underlying models to expand Claude's conversational abilities.
Read on TechCrunch AI →Anthropic has expanded voice mode access to its more capable Claude Opus and Sonnet models, moving beyond the previously exclusive Haiku model. The expansion includes integration with productivity apps like Gmail, Slack, and Canva, enabling users to tackle complex business problems through voice interaction.
Read on The Verge AI →Researchers introduced Athena-Brain-8B, an 8-billion parameter language model optimized as an on-device controller for robots. The model combines general language understanding with specialized embodied interaction capabilities through multi-stage training, achieving competitive performance on both general benchmarks and robot-specific tasks while producing concise responses for efficient hardware execution.
Read on arXiv cs.AI →Google released three new Gemini models (3.6 Flash, 3.5 Flash-Lite, and Flash Cyber) but notably did not release Gemini 3.5 Pro, prompting questions about the company's product roadmap and strategic direction.
Read on TechCrunch AI →Google DeepMind released three new Gemini models: version 3.6 Flash and two variants of version 3.5 (Flash-Lite and Flash Cyber). These additions expand the company's generative AI model lineup with different capability and performance configurations.
Read on Google DeepMind →Google DeepMind released new versions of its Gemini model family, including Gemini 3.6 Flash, 3.5 Flash-Lite, and a specialized 3.5 Flash Cyber variant. These updates expand the range of model sizes and capabilities available for different use cases.
Read on Google DeepMind →Google launched Gemini 3.5 Flash Cyber, a specialized AI model for identifying and fixing security vulnerabilities at lower cost than competing systems like Anthropic's Mythos. The model will initially be available to governments and trusted partners through Google's CodeMender coding agent.
Read on The Verge AI →Researchers introduced FUSAR-R1, a reasoning-focused vision-language model designed specifically for Synthetic Aperture Radar image interpretation. The model uses chain-of-thought reasoning data and reinforcement learning to perform expert-level analysis tasks like target detection, counting, and land-cover classification, outperforming existing multimodal models on SAR-specific challenges.
Read on arXiv cs.AI →Researchers introduced Pailitao-MMSearch, a specialized multimodal search model for e-commerce that handles text, images, and voice queries simultaneously. Built on Qwen and deployed on Taobao, the model achieved significant improvements in online metrics, including 13.61% increases in merchandise volume compared to traditional multimodal search approaches.
Read on arXiv cs.AI →Hugging Face announced Cosmos 3 Edge, a new model or tool in their Cosmos product line. The specific capabilities and technical details were not provided in the available information.
Read on Hugging Face →Chinese AI companies Moonshot and Alibaba released new models claiming competitive performance with leading U.S. systems like OpenAI and Anthropic, at lower costs. The releases signal narrowing technological gaps between Chinese and American AI capabilities amid growing geopolitical competition.
Read on The Verge AI →Researchers introduced Cura 1T, a specialized language model designed for healthcare tasks including patient consultation, clinical reasoning, and electronic health record tool use. The model was developed using a human-gated self-evolution training approach that iteratively improves capabilities by targeting observed failures with synthetic and curated data, achieving competitive performance on healthcare and general reasoning benchmarks.
Read on arXiv cs.AI →Researchers introduced S1-Omni, a multimodal AI model designed to handle diverse scientific tasks including molecular generation, protein structure prediction, and spectrum analysis. The unified model integrates scientific laws and expert knowledge while processing multiple data types, demonstrating competitive performance against specialized domain models and general-purpose AI systems on 60+ scientific benchmarks.
Read on arXiv cs.AI →Researchers introduced Loopie, a series of looped Transformer models using Mixture-of-Experts architecture that achieves competitive performance compared to larger vanilla Transformers trained with equivalent compute resources. The models demonstrate strong reasoning capabilities, achieving gold-medal results on the 2025 IMO and IPhO competitions.
Read on arXiv cs.AI →Moonshot AI, a Chinese company, released an updated version of its Kimi model, sparking debate about potential implications for AI development and geopolitical considerations.
Read on TechCrunch AI →Google unveiled Gemini 3.5 Flash Cyber, a specialized lightweight AI model designed to identify and remediate security vulnerabilities. The model combines efficiency with cybersecurity capabilities for vulnerability detection and patching.
Read on Google DeepMind →xAI's Grok 4.3 model is now available on Amazon Bedrock, enabling enterprises to build AI agents and workflows with a model featuring configurable reasoning, tool-use capabilities, and a 1 million token context window. The integration positions xAI as a new model provider on Bedrock's inference platform.
Read on AWS Machine Learning →NVIDIA released Nemotron 3 Embed, an embedding model that achieved the top ranking on the Retrieval Text Embedding Benchmark (RTEB). The model is designed to improve retrieval performance for agentic systems that require accurate information extraction.
Read on Hugging Face →Researchers introduced EZSMTV3, an enhanced framework for Constraint Answer Set Programming that combines logic programming with constraint solving. The system leverages existing SMT solvers and demonstrates improved language expressiveness, optimization capabilities, and support for mixed-domain constraints compared to competing CASP systems.
Read on arXiv cs.AI →Researchers released Boogu-Image-0.1, an open-source multimodal model family that performs text-to-image generation, editing, and bilingual text rendering. The model achieves performance comparable to closed-source alternatives while trained on 208.62 million images with approximately $400K computational cost, with code and weights made publicly available.
Read on arXiv cs.AI →OvisOCR2 is a 0.8B document parsing model that converts document images to Markdown format covering text, formulas, tables, and visual elements. Trained using a combination of real and synthetic data with supervised fine-tuning and reinforcement learning, it achieves state-of-the-art performance on multiple benchmarks, surpassing pipeline-based methods in end-to-end document parsing.
Read on arXiv cs.AI →Thinking Machines Lab released Inkling, an open-source model with 975 billion parameters designed to process both video and audio. The release positions the startup as a competitor in the generative AI space alongside established players like Anthropic and OpenAI.
Read on Wired AI →Thinking Machines released Inkling, its first open-source AI model, marking the company's public debut after 18 months of developing AI infrastructure privately. The release positions the company against the trend of universal large language models, instead offering a specialized alternative.
Read on TechCrunch AI →Alibaba researchers introduced QwenPaw-Data, an autonomous agent system designed for enterprise data analysis that integrates data warehouses, dashboards, and documents to convert natural language requests into analytical workflows. The system uses interconnected metadata graphs, reusable analytical skills, and controlled execution to improve data access reliability and analytical quality while continuously learning from feedback.
Read on arXiv cs.AI →Apple has redesigned Siri to function as a central hub for iPhone functionality rather than just a voice assistant. The updated version is available for testing through the iOS 27 public beta.
Read on Wired AI →OpenAI's GPT-5.6 model family (Sol, Terra, and Luna) is now available through Amazon Bedrock's inference platform. The models are designed for enterprise workloads requiring extended reasoning and high reliability, with pricing aligned to OpenAI's direct rates.
Read on AWS Machine Learning →Apple released the first public beta of iOS 27, which prioritizes performance improvements and bug fixes over new features. Key updates include faster app launches and search, enhanced Messages functionality with RCS encryption, and refinements to Liquid Glass technology.
Read on The Verge AI →Researchers released Soofi S 30B-A3B, an open-source hybrid language model optimized for German and English that uses a Mixture-of-Experts architecture to activate only 3B of its 30B parameters per token. Trained on 27 trillion tokens with emphasis on German, it matches larger dense models on benchmarks while achieving superior code performance and outperforming other European sovereign models, with all weights and training artifacts to be publicly released.
Read on arXiv cs.AI →Researchers introduced ALICE, a unified foundation model for pathology that combines knowledge from eight specialized models through multi-stage distillation. Trained on nearly 25 million pathology images, ALICE outperformed task-specific models across 21 evaluation scenarios and 96 downstream tasks in tissue analysis, vision-language tasks, and whole-slide assessment.
Read on arXiv cs.AI →Researchers introduced Infinity-Parser2, a multimodal model for document parsing that uses synthetic data generation and multi-task reinforcement learning. The team released a 5-million-sample bilingual dataset and two model variants—Flash for speed and Pro for accuracy—achieving state-of-the-art results on multiple parsing benchmarks.
Read on arXiv cs.AI →OpenAI has released a new model family that includes GPT-5.6, offering enhancements in multiple domains with a particular focus on cybersecurity capabilities.
Read on TechCrunch AI →Meta launched Muse Spark 1.1, an AI coding assistant designed to tackle large-scale enterprise tasks like bug fixing and code migrations. The tool targets businesses seeking automation capabilities for complex development workflows.
Read on TechCrunch AI →OpenAI released GPT-5.6 to the public following regulatory approval from the Trump administration, after an initial limited preview period. The company simultaneously launched ChatGPT Work, a new agent combining ChatGPT with code-like capabilities for non-technical users, powered by GPT-5.6.
Read on The Verge AI →Meta released Muse Spark 1.1, an upgraded coding model accessible through its new Model API for developers. The updated version offers improved bug detection, better support for multi-agent workflows, and multimodal capabilities across images, videos, and documents.
Read on The Verge AI →OpenAI announced GPT-5.6, a new model offering improved efficiency and performance metrics. The model provides enhanced computational output per token and reduced cost-per-performance while delivering increased capability for complex tasks.
Read on OpenAI →OpenAI announced ChatGPT Work, an agent designed to execute tasks across multiple applications and files while maintaining context over extended sessions to complete complex projects.
Read on OpenAI →Microsoft 365 Copilot now uses GPT-5.6 as its underlying model, providing enhanced AI capabilities across productivity applications including Word, Excel, PowerPoint, and Chat tools.
Read on OpenAI →SpaceX's AI division released Grok 4.5, positioning it as a cost-effective competitor to other advanced AI models. Elon Musk characterized it as comparable in capability to high-performance alternatives while offering improved efficiency.
Read on TechCrunch AI →OpenAI introduced new voice models enabling simultaneous speaking and listening capabilities, which the company describes as important for real-time translation applications.
Read on TechCrunch AI →OpenAI unveiled GPT-Live-1, an upgraded voice model for ChatGPT that interrupts users less frequently and better simulates natural conversation. The model can delegate tasks to more capable text models like GPT-5.5 for complex reasoning or web searches, improving response quality and speed.
Read on The Verge AI →KAT-Coder-V2.5 is an autonomous coding agent trained to work directly within real code repositories rather than generating isolated snippets. The model employs advanced post-training techniques including sandboxed environment reconstruction, reinforcement learning optimization, and multi-teacher distillation, achieving competitive performance on software engineering benchmarks.
Read on arXiv cs.AI →Harrison.Rad 1.5 is a radiology-focused AI model that generates medical reports by analyzing X-ray images alongside patient history and clinical context. Trained through domain adaptation and vision-language techniques, it achieved the highest performance on clinical benchmarks and meets standards equivalent to professional radiology certification exams.
Read on arXiv cs.AI →OpenAI has launched GPT-Live, an updated voice model generation designed to enable more natural conversations between humans and AI systems. The technology now powers ChatGPT's voice feature.
Read on OpenAI →Meta has launched Muse, an AI image generation model designed for applications across advertising, interior design, and content creation. The tool expands Meta's offerings in generative AI capabilities.
Read on TechCrunch AI →Meta launched Muse Image, an AI image generation model from its Superintelligence Labs, now integrated across Meta AI, Instagram, and WhatsApp with planned expansion to Facebook and Messenger. The model can reason through prompts and search the web before generating images, representing part of Meta's transition from Llama to the Muse model family.
Read on The Verge AI →Anthropic is expanding Claude Cowork, its collaborative AI platform, to mobile and web interfaces starting Tuesday. Previously limited to desktop apps, the feature will roll out first to Max subscribers, with full functionality remaining on desktop while mobile and web versions offer core collaboration capabilities.
Read on The Verge AI →Anthropic has extended its Claude Cowork agent to mobile phones, allowing it to continue executing tasks after the user closes their laptop. This represents the company's broader strategy to develop smartphone-accessible AI agents.
Read on Wired AI →Google announced expanded capabilities for Managed Agents in the Gemini API, enabling developers to build production-grade agents with features including background task execution and remote model context protocol support.
Read on Google AI Blog →iFLYTEK released Embodied-Omni, a unified multimodal foundation model that integrates vision, language, and action processing in a single framework for robotic control tasks. The model uses a brain-cerebellum architecture where vision-language components handle high-level planning while action components directly execute instructions, trained on both human demonstrations and robot interaction data.
Read on arXiv cs.AI →Researchers introduced OpenDDE, an open-source AI model for drug discovery that predicts biomolecular structures and relationships using co-folding technology. The system integrates structure prediction with additional capabilities for drug design and optimization, with released code and benchmarks intended to advance collaborative research in computational biology.
Read on arXiv cs.AI →Nvidia researchers introduced Nemotron-Labs-3-Puzzle-75B-A9B, a compressed version of their Nemotron-3-Super model designed for efficient deployment. Using a multi-stage compression pipeline combining pruning, distillation, and quantization, the model achieves 2x higher server throughput on interactive workloads and enables 8x more concurrent long-context requests on a single H100 GPU while maintaining competitive performance across reasoning, coding, and multilingual benchmarks.
Read on arXiv cs.AI →Google introduced Gemma 4, an open-weight multimodal language model family ranging from 2.3B to 31B parameters. The models feature improved vision and audio processing, a reasoning mode for generating explanations, and optimizations for inference efficiency and long-context understanding, achieving performance competitive with larger models on various benchmarks.
Read on arXiv cs.AI →AWS introduced Reverse Direct Preference Optimization (rDPO), a technique that allows organizations to selectively reduce content moderation restrictions in Amazon Nova models for legitimate business purposes. The Customizable Content Moderation Settings feature lets approved customers adjust safeguards across safety, and other responsible AI categories while maintaining model quality.
Read on AWS Machine Learning →Researchers introduced Wiola, a novel small language model architecture featuring five new components: spiral rotary positional encoding, gated cross-layer attention, adaptive token merging, dual stream feed-forward networks, and modified normalization. The architecture was released in four sizes ranging from 120M to 1.5B parameters and is compatible with HuggingFace Transformers.
Read on arXiv cs.AI →Researchers developed FitOne, a series of specialized language models (8B and 32B parameters) optimized for fitness coaching by applying domain-specific post-training techniques. The models demonstrate 7-13% improvements over general-purpose baselines on professional fitness certification exams while maintaining broad capabilities.
Read on arXiv cs.AI →Seed2.0 is a new model series designed to handle complex real-world tasks by addressing long-tail knowledge gaps and improving instruction-following capabilities. The developers built a specialized evaluation framework based on genuine user needs and realistic scenarios to guide the model's development, achieving improvements in reasoning, visual understanding, and search functionality.
Read on arXiv cs.AI →Amazon Bedrock now offers OpenAI's GPT OSS models and NVIDIA's Nemotron models in AWS GovCloud, enabling U.S. government agencies to access advanced open-weight models while maintaining security and compliance requirements without moving sensitive data outside their boundaries.
Read on AWS Machine Learning →Google has released Gemini Spark, an agentic assistant designed to operate continuously, on macOS platforms. The expansion includes new capabilities such as real-time tracking and broader app integration.
Read on TechCrunch AI →Xiaomi introduced Xiaomi-GUI-0, a GUI agent trained and evaluated on real mobile devices rather than simulated environments, addressing the gap between benchmark performance and real-world usability. The system uses a hybrid infrastructure combining physical devices with sandboxes and employs multi-source training data with an error-driven feedback loop to improve stability across real applications.
Read on arXiv cs.AI →Cohere and LG CNS developed LuckyStar 111B, a multilingual model optimized for Korean-English enterprise agents. The model uses supervised fine-tuning, reinforcement learning, and quantization techniques to enable efficient tool-use and reasoning while maintaining performance under memory constraints.
Read on arXiv cs.AI →AWS announced that Anthropic's Claude Fable 5 models will be available on Amazon Bedrock with enhanced safety features. AWS emphasized its commitment to secure model deployment while balancing rapid access to frontier AI capabilities for customers against broader societal security considerations.
Read on AWS Machine Learning →Google released an updated version of its image generation model that offers faster processing speeds and reduced costs. The improvements aim to make AI image creation more accessible and practical for content creators.
Read on TechCrunch AI →Anthropic released Claude Sonnet 5 on Amazon Bedrock and Claude Platform on AWS, positioning it as the most capable Sonnet model with improved performance for coding and agentic tasks while maintaining cost efficiency. The model integrates with AWS infrastructure, offering enterprise security, regional data residency, and unified billing alongside Anthropic's native platform features.
Read on AWS Machine Learning →Anthropic released Claude Sonnet 5, a new model designed to run AI agents more affordably while offering improved agentic abilities and safety features. The model positions itself as a cost-effective alternative to premium offerings from competitors.
Read on TechCrunch AI →Google DeepMind announced new AI models available for development: Nano Banana 2 Lite and Gemini Omni Flash. These models appear designed to offer developers options for building applications with different performance and capability tradeoffs.
Read on Google DeepMind →Base44, a vibe coding platform owned by Wix, is launching its own AI model to differentiate itself from competitors and potentially surpass leading frontier models. The move reflects a trend where AI startups develop proprietary models to establish competitive advantages and reduce dependency on third-party providers.
Read on TechCrunch AI →Google has made Gemini's personalized image generation feature available to eligible free users in the United States. The feature leverages user interests and data from connected Google apps to generate customized images.
Read on TechCrunch AI →Zhipu AI released GLM-5.2, an open-weight model that researchers claim matches performance with advanced security-focused models in bug-finding and cybersecurity tasks. Though the model trails US competitors on general benchmarks, the release signals China's narrowing capability gap, raising concerns for the US government's ongoing efforts to restrict China's access to advanced AI systems.
Read on The Verge AI →Asian AI startups are releasing models with capabilities comparable to Anthropic's offerings, positioning them as alternatives to U.S. products amid concerns about American export restrictions. This development highlights growing competition in non-U.S. AI markets as regional companies fill potential gaps left by export limitations.
Read on TechCrunch AI →Researchers introduced ReasonCLIP-58M, an enhanced version of CLIP that incorporates reasoning supervision through a two-stage training approach. The framework includes new datasets and a benchmark to improve visual reasoning capabilities, and shows performance gains when integrated into multimodal systems like LLaVA-NeXT without increasing inference costs.
Read on arXiv cs.AI →Researchers introduced Wan-Streamer, a foundation model built for real-time audio-visual interaction with approximately 200ms response latency. Unlike traditional systems using separate modules for speech recognition, language processing, and video generation, this model integrates all capabilities in a single Transformer architecture with specialized streaming mechanisms.
Read on arXiv cs.AI →The Trump administration has authorized over 100 US companies and government agencies to use Anthropic's Mythos 5 model, with access extended to their international employees. This marks a significant expansion of model availability across corporate and federal sectors.
Read on TechCrunch AI →OpenAI released GPT-5.6, a new model suite with three variants (Sol, Terra, Luna) following a staggered release arrangement requested by the Trump administration. The flagship Sol model emphasizes capabilities in coding, cybersecurity, and biology, with pricing at $5/$30 per million tokens.
Read on The Verge AI →OpenAI is developing Jalapeño, a custom inference chip created with Broadcom, to reduce reliance on Nvidia's dominance in AI hardware. This move joins other major tech companies in building proprietary chips to diversify their supplier dependencies.
Read on TechCrunch AI →OpenAI announced GPT-5.6 Sol, a new model featuring improved performance in coding, scientific reasoning, and cybersecurity applications, accompanied by enhanced safety measures.
Read on OpenAI →Researchers released Qwen3-Instruct SAE, a collection of sparse autoencoders trained on Qwen3 language models to extract interpretable features from neural representations. The work demonstrates how these SAEs can identify and manipulate specific behaviors, such as refusal responses, providing tools for understanding and steering model behavior.
Read on arXiv cs.AI →ZONOS2 8B, a new text-to-speech model, scales up from its predecessor to 8 billion parameters using a mixture-of-experts architecture and trains on over 6 million hours of audio data. The model achieves competitive performance on naturalness, prosody, and voice cloning while maintaining efficient streaming latency, with weights and code released publicly.
Read on arXiv cs.AI →Researchers introduced FISHER, a foundation model designed for analyzing industrial signals across multiple data types and sampling rates. The model uses a novel sub-band approach to handle varying sampling rates without resampling and is pre-trained using self-distillation on audio data. FISHER outperforms larger specialized models while being significantly smaller, with researchers releasing both the model and a new 19-dataset benchmark.
Read on arXiv cs.AI →Google DeepMind has added computer use capabilities to Gemini 3.5 Flash, enabling the model to interact with digital interfaces and perform tasks directly on computers rather than just providing instructions.
Read on Google DeepMind →OpenAI unveiled Jalapeño, a custom processor developed by Broadcom designed to optimize inference operations for OpenAI's AI systems. This move represents the company's efforts to build specialized hardware infrastructure to improve computational efficiency.
Read on TechCrunch AI →OpenAI unveiled Jalapeño, a custom AI processor chip developed with Broadcom designed specifically for AI inference tasks in servers. The ASIC chip is intended to support both current and future large language models, marking OpenAI's entry into custom hardware for AI operations.
Read on The Verge AI →OpenAI and Broadcom jointly developed Jalapeño, a specialized chip designed to optimize large language model inference workloads with improved performance and energy efficiency.
Read on OpenAI →OpenAI announced an enhanced version of GPT-5.5-Cyber and launched the 'Patch the Planet' initiative aimed at identifying and fixing vulnerabilities in open-source software, responding to growing concerns about AI security capabilities.
Read on Wired AI →PaddlePaddle released PP-OCRv6, an optical character recognition model supporting 50 languages with varying parameter sizes from 1.5M to 34.5M. The model is now available on Hugging Face, offering options for different computational requirements.
Read on Hugging Face →Apple is introducing multiple AI capabilities across iOS 27 beyond Siri improvements, with practical features being integrated into various parts of the operating system. These additions were demonstrated alongside Siri's redesign at WWDC.
Read on TechCrunch AI →Apple's updated Siri AI features improved conversational abilities and expanded availability across devices, with enhanced practical utility for users. The assistant demonstrates stronger natural language understanding and responsiveness compared to previous iterations.
Read on Wired AI →Researchers introduced IHUBERT, a Persian language model trained on 7-8 billion tokens using semantic deduplication and domain-balanced preprocessing to address data scarcity. The 125M-parameter model achieved top performance on Persian question answering and natural language inference tasks while remaining competitive on other NLU benchmarks.
Read on arXiv cs.AI →Researchers introduced SleepMaMi, a foundation model designed for sleep medicine that analyzes both full-night sleep patterns and detailed biosignal features from polysomnography recordings. Trained on over 20,000 PSG recordings, the model uses dual encoders and demographic-guided learning to outperform existing approaches across multiple clinical sleep analysis tasks.
Read on arXiv cs.AI →TerraMind is a new multimodal foundation model designed for Earth observation that processes data at both token and pixel levels across nine geospatial data types. The model demonstrates strong performance on standard benchmarks and introduces a capability to generate synthetic data during training and inference, with researchers releasing the model weights and code openly.
Read on arXiv cs.AI →Researchers introduced Vero, an open-source family of vision-language models trained with reinforcement learning across diverse visual reasoning tasks. The framework combines 600K samples from 59 datasets with task-specific reward mechanisms, enabling models to match or exceed existing open-weight competitors on benchmarks spanning charts, science, and spatial understanding.
Read on arXiv cs.AI →DeepSeek released V4 series language models featuring two MoE variants (Pro with 1.6T parameters and Flash with 284B parameters) that support one million token contexts. The models incorporate architectural improvements including hybrid attention mechanisms and achieve significantly improved efficiency, requiring 27% fewer inference FLOPs and 10% of KV cache compared to the previous V3.2 version.
Read on arXiv cs.AI →
Hi — I'm Scott. I'm building AI Info Forge because the Claude skills market is full of prompts that worked last month and break this month. Verified, drift-tested skills is the version of this market I wish existed. If you've felt the same friction, get on the list and I'll send you early access.
— Scott Sokolowski, Founder