China’s Seedream 4.0 Just Outperformed Google’s Nano Banana (And It’s Open Source)

ByteDance's new image generation model, Seadream 4.0, claims to outperform Google's Nano Banana, marking China's strong entry into multimodal AI, while other tech giants expand their AI initiatives in dating, visual reasoning, and infrastructure.

• ByteDance's Seadream 4.0 challenges Google's Nano Banana with advanced image generation and editing capabilities.

• Meta introduces AI dating features on Facebook, while Alibaba launches open-source visual language models with deep reasoning.

• OpenAI, Oracle, and SoftBank are investing hundreds of billions in a massive expansion of AI data centers across the U.S.

ByteDance has released Seadream 4.0, an AI image generation model that claims to surpass Google DeepMind's Nano Banana in key areas like prompt accuracy and visual quality, showcasing China's increasing competition in AI. Meanwhile, Meta has introduced AI-powered dating features in Facebook, and Alibaba launched advanced open-source visual language models focusing on reasoning and execution over images and video. Massive investments in AI data centers are also underway by OpenAI, Oracle, and SoftBank.

ByteDance Seadream 4.0 Launch

• 00:00:32 ByteDance, the company behind TikTok, released Seadream 4.0, an AI-powered image generation model, openly claiming it outperforms Google DeepMind's Gemini 2.5 flash image, known as Nano Banana. Seadream 4.0 combines text-to-image creation with editing capabilities, featuring a new architecture that is 10 times faster in raw image inference than previous versions. Internal benchmarks on MagicBench show Seadream 4.0 scored higher than Gemini 2.5 Flash image in prompt accuracy, alignment, and overall visual quality, though a technical report is yet to be published.

Seadream 4.0 Capabilities and Access

• 00:02:16 Seadream 4.0 offers new capabilities, including image generation in 2K and 4K resolution, the ability to feed up to six reference images, and support for any aspect ratio. Early feedback highlights its accurate quick changes to existing images using text prompts. While priced at $30 per 1000 generations, positioning it as a professional-grade service, it is live in China via Jemang and Dubao AI apps, and globally through a partnership with FreePic, indicating China's aggressive push into image generation AI.

Meta's AI Dating Features

• 00:03:18 Meta has updated Facebook Dating with two AI-powered features, a Dating Assistant and MeetCute, aiming to foster deeper, more authentic connections beyond traditional swiping. The AI Dating Assistant refines profiles, suggests matches, and offers personalized recommendations based on niche prompts like 'find me a Brooklyn girl in finance.' MeetCute provides one unexpected match weekly to break users out of their usual selection loops, initially rolling out in the U.S. and Canada for the 18-29 age group amidst Match Group's financial struggles.

Alibaba's Advanced Visual AI

• 00:05:27 Alibaba's Quen team unveiled Quen3VL, its most advanced vision language model, including the open-source flagship model QUEN3VL235BA22B, available in INSTRUCT and THINKING versions. These models excel in reasoning and execution over visual data, moving beyond simple object recognition to break down relationships, handle complex visual math, and interact with digital environments. With native support for 256,000 tokens of context, expandable to 1 million, and innovations like DeepStack technology, Quen3VL claims to match or surpass Gemini 2.5 Pro in visual perception and outperform it in math-heavy challenges, solidifying China's role in multimodal AI.

Massive AI Data Center Expansion

• 00:09:09 OpenAI, Oracle, and SoftBank announced a significant expansion of Stargate, their mega AI data center project in the United States, adding five new sites and pushing planned capacity to nearly seven gigawatts with over $400 billion investment for the next three years. The goal is 10 gigawatts and $500 billion by the end of 2025, with development already ahead of schedule. These builds are expected to create more than 25,000 on-site jobs and tens of thousands more indirectly across sites in Texas, New Mexico, Ohio, and an unannounced Midwest location.