This week in AI, creativity, visual generation, and robotics take a massive leap forward. From tools that let anyone remix videos to breakthroughs powering smarter robots, here’s what’s new:

🎥 Meta Vibes
A new feed of endlessly scrollable AI videos you can create, remix, and share directly to Instagram or Facebook, making video creation fast, fun, and personalized.

🖼️ Hunyuan Image 3.0
Tencent’s largest open-source text-to-image model delivers photorealistic and artistic visuals with accurate text rendering, advanced multimodal architecture, and prompt refinement, empowering creators and businesses worldwide.

🤖 Gemini Robotics 1.5
DeepMind’s latest model gives robots real-world adaptability, combining vision, language, and action to perform tasks, use tools, and respond naturally.

From creativity to capability, these innovations show how AI is expanding its reach into how we create, build, and interact with the world.

Meta Vibes: Remix and Create AI Videos Easily

Meta has launched Vibes, a new feature in the Meta AI app and on meta.ai that offers a feed of endlessly scrollable, short-form AI-generated videos, which users can create, remix, and share directly or cross-post to Instagram Reels and Facebook Stories. The feed aims to inspire creativity by letting anyone start from scratch, use their own media, or remix videos seen in the feed, with the ability to add visuals, music, and adjust styles for personalization. Over time, Vibes will become more personalized, and Meta is working with artists and creators to roll out powerful new tools and models, positioning the platform as both a creative playground and a showcase for AI-generated content in over 40 countries (excluding the UK).

Hunyuan Image 3.0: Tencent’s Next-Gen AI Image Creator

HunyuanImage 3.0, Tencent’s largest open-source text-to-image model, has 80 billion parameters with 13 billion activated per token during inference. Built on a native multimodal large language model, it excels in world knowledge reasoning, ultra-long prompt understanding, and precise text generation within images. Its unique Mixture of Experts (MoE) and Transfusion architecture deeply couples diffusion with LLM training for seamless multimodal integration. Trained on 5 billion image-text pairs and 6 trillion text tokens, HunyuanImage 3.0 delivers photorealistic and artistic images with advanced prompt enhancement and refiner modules. This model dramatically accelerates creative workflows for illustrators, designers, and creators by generating intricate comics, emojis, and educational visuals. It supports future updates like image editing and multi-turn interaction while offering commercial use with open-source availability.

DeepMind’s Breakthrough in Smarter, Adaptive Robots

Gemini Robotics 1.5 by Google DeepMind is a cutting-edge AI system that helps robots of all shapes and sizes see, understand, and interact with the world around them. It can plan and carry out complex tasks by breaking them down into simple steps, think before acting, and explain what it's doing in plain language. This system learns skills from one robot and applies them to others, speeding up robot learning. It can also be guided with everyday language, making it easy for people to work alongside robots. Supporting everything from simple robotic arms to humanoid robots, Gemini Robotics 1.5 is designed to make robots smarter, safer, and more helpful in daily life.

Hand Picked Video

In this video, we’ll look at how Google’s Gemini Nano Banana model edits complex images, replacing objects, improving vibes, restoring photos, and even making creative transformations.

Top AI Products from this week

• Integrity - Integrity is the unified project brain where notes, canvases, and AI chats become connected layers of thought. Move seamlessly from structure to vision, and co-create with the top AI models, grounded in your work.

• Deamoy – Inspired by Figma’s craft and Lovable’s magic, Deamoy fuses the best of both: streamable JIT rendering, AI and visual edits in perfect sync, and editable sites from a single prompt in seconds. The fastest, freest AI app builder is here.

• Nexa SDK – Nexa SDK runs any model on any device, across any backend locally—text, vision, audio, speech, or image generation—on NPU, GPU, or CPU. It supports Qualcomm and Apple NPUs, GGUF, Apple MLX, and the latest SOTA models (Gemma3n, PaddleOCR).

• opencode zen - opencode zen is a curated, “ready-to-use” gateway of tested models optimized for coding agents. It integrates as an optional provider within opencode, offering pay-as-you-go access to top models (e.g. GPT-5, Claude, Grok) with no lock-in.

• Hoodiely - Create custom shirts by tweeting! Mention @HoodielyHQ in any tweet or reply and get a unique, personalized shirt design. High-quality printing, worldwide shipping. Pay only if you love it.

This week in AI

• Perplexity Search API Launch - Access hundreds of billions of webpages via API with real-time updates, precise sub-document snippets, SDK, open evaluation, and low-cost pricing. Perfect for AI apps.

• AI Smart Bandage Heals 25% Faster - New gene therapy AMT-130 slows Huntington’s progression by 75% over 3 years with one-time brain surgery, offering hope for the first effective treatment.

• AI-Made Psychedelic Without Trip - Mindstate Design Labs created "the least psychedelic psychedelic" using AI, offering psychoactive benefits without hallucinations, validated by successful Phase I trials.

• Toyota & Isuzu Develop Fuel Cell Bus - Toyota and Isuzu announced joint development of next-gen fuel cell route buses, with production starting FY2026, aiming to reduce CO2 emissions and advance carbon neutrality.

• GDPval AI vs Experts Real-World Tasks - OpenAI's GDPval benchmark tests AI on 1,320 real expert tasks across 44 US occupations, showing AI nearing expert-level quality, faster and cheaper on routine jobs.Paper of The Day

Paper of The Day

Dynamic Experts Search: Enhancing Reasoning in Mixture-of-Experts LLMs at Test Time proposes a new Test-Time Scaling (TTS) strategy for large language models using Mixture-of-Experts (MoE) architectures. Unlike traditional methods that only diversify outputs via sampling, DES dynamically varies the number of activated experts during inference to explore diverse reasoning paths. By maintaining consistent expert activation within each reasoning trajectory, DES allocates computation effectively, improving accuracy and reasoning stability without extra cost. Experiments across math, code, and knowledge tasks show DES outperforms existing TTS baselines, offering a scalable, architecture-aware method to enhance LLM reasoning capabilities.

To read the whole paper 👉️ here.