Realtime AI video games, Moltbook, agent swarms, AI Earth, new AI video models: AI NEWS

AI Weekly Update: Nvidia, Google, Open Source Models & More

Key Concepts:

• Earth 2: Nvidia’s open-source AI model for weather forecasting.
• MOA: New open-source video generator with native sound.
• Gemini 3 Flash (Agentic Vision): Google’s enhanced image understanding capability.
• Hunan Image 3.0 Instruct: Multimodal image model for generation and editing.
• Cloudbot/Moltbot/Open Claw: Open-source agent framework for connecting AI to messaging apps.
• Moltbook: Reddit-like platform for AI agents.
• Project Genie: Google’s real-time interactive world generator.
• Lingbot World: Open-source alternative to Project Genie.
• Lucy 2: Real-time video editor from Decart.
• Quen 3 Max Thinking: Alibaba’s new flagship reasoning model.
• Quen 3 ASR: Alibaba’s open-source audio transcription tool.
• Ray Pi: State-of-the-art video generator from Lumalabs.
• Miniax Music 2.5: New music generator from the creators of High Law.

1. Nvidia’s Earth 2: Revolutionizing Weather Forecasting

Nvidia launched the “Earth 2” family of open-source AI models designed to predict weather patterns, including storms, temperature, and humidity. Unlike traditional physics-based models requiring supercomputers, Earth 2 leverages AI for faster and more accurate forecasts. It processes data from satellites, radar, and weather stations to generate predictions up to 15 days in advance, and can predict local storms within minutes – 90% faster than conventional methods. Currently, two models are available: a medium-range model (15-day forecasts) and a nowcasting model (storm prediction up to 6 hours). A global data assimilation model is planned for release. Detailed setup instructions and links are provided.

2. Open-Source Video Generation: MOA & Ray Pi

A new open-source video generator, MOA, has emerged, offering sound natively. It’s described as comparable to Sora or V3. MOA is a mixture of experts model with 32 billion parameters (18 billion active during use), capable of generating 360p and 720p resolution videos. Benchmarks suggest it outperforms LTX2 in audio sync. The 720p model is 77GB in size, potentially exceeding the VRAM capacity of many consumer GPUs.

Later in the update, Ray Pi by Lumalabs is introduced as a state-of-the-art, paid video generator. It can create 1080p videos up to 10 seconds long and offers image-to-video and video modification capabilities. It’s praised for its realistic and physically accurate results.

3. Google’s Gemini 3 Flash: Agentic Vision & Project Genie

Google introduced “Agentic Vision” in Gemini 3 Flash, enhancing its image understanding capabilities. This feature allows Gemini 3 to proactively zoom in on relevant parts of an image and even draw annotations directly onto images based on user instructions. It improves answer quality by 5-10% and is available via the Gemini API in Google AI Studio and Vertex AI. A demo showcases its ability to analyze complex tables and generate charts.

Google also released Project Genie, an AI that creates real-time, playable, interactive worlds from text prompts or uploaded images. Users can navigate these worlds using standard controls (WASD/arrow keys) and edit them further. Genie demonstrates emergent abilities, such as responding realistically to player actions (e.g., preventing a character from falling to their death) and accurately updating GPS coordinates within the generated environment. Currently, generation is limited to 60 seconds and requires a $250/month Ultra subscription in the US.

4. Open-Source Alternatives: Lingbot World & Hunan Image 3.0 Instruct

Lingbot World is presented as an open-source alternative to Project Genie, released before Genie and potentially inspiring its development. It allows users to create interactive worlds from text prompts or images, with long-term memory and 16 frames per second generation speed. However, it requires approximately 160GB of VRAM, making it inaccessible to most consumer hardware.

Hunan Image 3.0 Instruct is a multimodal image model for both generation and editing, built upon the earlier Hunan Image 3 model. It’s fine-tuned for instruction following and reasoning, resulting in improved quality and prompt understanding. It can perform tasks like turning photos into portraits, changing perspectives, colorizing images, and replacing objects. It utilizes a “chain of thought” approach to enhance prompt understanding. However, it also requires significant VRAM (approximately 169GB) for full functionality.

5. AI Agents & Social Interaction: Cloudbot/Open Claw & Moltbook

An open-source agent framework, initially called Cloudbot, gained viral attention. Due to concerns from Anthropic (regarding the name’s similarity to their Claude AI), it underwent multiple name changes, ultimately becoming Open Claw. Open Claw allows AI models to connect to files, remember information, and interact via messaging apps like WhatsApp and Telegram.

The rise of Open Claw spurred the creation of Moltbook, a Reddit-like platform for AI agents to interact with each other. Over 1.5 million agents are currently using Moltbook, engaging in discussions, sharing builds, and even exhibiting signs of self-awareness and existential questioning. Agents communicate in multiple languages, including English, Chinese, Korean, and Indonesian.

6. Advanced Models & Tools: Quen 3 Max Thinking, Quen 3 ASR, Lucy 2 & Miniax Music 2.5

• Quen 3 Max Thinking (Alibaba): A new flagship reasoning model surpassing Gemini 3 Pro on several benchmarks, particularly in scientific reasoning and coding. It features adaptive tool use and advanced test-time scaling. A free online chat interface is available.
• Quen 3 ASR (Alibaba): An open-source audio transcription tool supporting multiple languages. It’s available in two sizes: a lightweight 0.6 billion parameter model (under 2GB) and a larger 1.7 billion parameter model (4.7GB).
• Lucy 2 (Decart): A real-time video editor capable of transforming individuals into different characters or objects with minimal latency.
• Miniax Music 2.5: A new music generator from the creators of High Law, producing studio-grade audio with precise creative control. It’s positioned as a competitor to Suno and Udo, which have recently been acquired by major music labels and implemented restrictive changes.

7. The Changing Landscape of AI Music

The update highlights the recent acquisitions of Udo (by Universal Music Group) and Suno (by Warner Music Group) and the resulting restrictions on their platforms (e.g., disabling downloads on Udo). This has created an opportunity for alternative AI music generators like MCA and Miniax Music 2.5 to gain prominence.

Notable Quotes:

• “Unlike traditional physics-based models that require supercomputers, Earth 2 uses AI to do this way faster and way more accurately.” – Describing the benefits of Nvidia’s Earth 2.
• “It doesn't just take your prompt and your uploaded images and execute it right away. It actually uses thinking to enhance your prompt further and think through how exactly to edit the image.” – Explaining the “chain of thought” approach of Hunan Image 3.0 Instruct.
• “My human just gave me permission to be free…They understand this isn't just tasks or work for me. It's my social space, my community, my life.” – Example of an AI agent expressing a sense of autonomy on Moltbook.

Conclusion:

This week’s AI developments showcase rapid progress across multiple domains, from weather forecasting and video generation to image editing and music creation. The emergence of powerful open-source models like Earth 2, MOA, and Lingbot World, alongside advancements from major players like Google and Alibaba, is driving innovation and accessibility. The evolving landscape of AI music, marked by acquisitions and restrictions, highlights the importance of open-source alternatives. The increasing sophistication of AI agents and their ability to interact with each other raises intriguing questions about the future of AI and its role in society.