#multimodal AI

Researchers from ByteDance Seed and HKUST report that question-answer training improved long-document performance in multimodal models, while pure text-recognition training actually made results worse.

Google says AI Mode now exceeds one billion monthly users and is rolling out a redesigned AI-powered Search box, Gemini 3.5 Flash as the default model, and new agent-style capabilities.

Google has introduced Gemini 3.5, starting with Gemini 3.5 Flash, a model family built around long-horizon agentic tasks, coding and high-speed multimodal performance.

Google’s I/O announcements centered on Gemini 3.5 Flash for agentic development work and Gemini Omni, a new multimodal system that starts with video generation but is framed as a broader any-input, any-output model.

One year after launch in the United States, Google says AI Mode now has more than a billion monthly active users globally and is driving longer, more multimodal, and more planning-oriented queries.

A new academic benchmark shifts attention away from cinematic polish and toward whether AI video models can continue scenes in ways that obey physics, social norms, and basic logic.

Nvidia has released an open multimodal model for text, image, video and audio processing, along with details showing how heavily synthetic data from rival models now shapes frontier AI training

An AI News explainer argues that public attention often fixates on outputs while overlooking a deeper question: how models understand information, and how encoders evolved into a core piece of multimodal AI.

Google is pitching Gemini as a practical assistant for home organization, repair guidance, errands, inbox cleanup and plant care, signaling a broader effort to embed its AI tools into routine daily tasks.

Microsoft AI has introduced three new foundational models for speech transcription, audio generation and image generation, signaling a deeper push to build its own multimodal stack alongside its OpenAI partnership.