This week in AI has been exceptionally busy with multiple groundbreaking releases across video generation, 3D modeling, and open-source AI tools. Tencent’s Hunyan World 1.5 introduces a real-time interactive 3D world generator that allows users to explore dynamically created scenes with impressive quality and responsiveness. Unlike traditional pre-built game worlds, this AI generates environments and characters on the fly, supporting first-person and third-person views, and even allows environmental changes like smoke or explosions through simple prompts. The model is open-source and can run locally with a CUDA GPU of at least 14 GB VRAM, making it accessible for developers interested in real-time 3D content creation.
Another notable innovation is Stereo Space, an AI that converts 2D images into 3D stereo images viewable with traditional 3D glasses or by crossing your eyes. This model outperforms other 3D photo generators in benchmarks and is available as a free Hugging Face space for users to try. Meanwhile, Long V2 pushes the boundaries of AI video generation by enabling ultra-long videos up to five minutes with consistent quality and coherence, a significant improvement over previous models limited to around 10 seconds. This model also requires around 14 GB VRAM and is open-source, allowing users to generate longer AI videos locally.
In character animation, the new Scale AI model stands out as the best open-source tool for full-body control and complex movements. It extracts 3D poses from reference videos and applies them accurately to any character, regardless of shape or proportions, outperforming competitors like One Animate and Vase. This tool supports multi-character animation and is integrated into Comfy UI, making it highly practical for animators seeking precise and consistent motion transfer. Additionally, Alibaba’s Quen Image Layered model offers advanced image editing by decomposing images into editable layers, similar to Photoshop, enabling targeted edits without affecting the entire image. Though large in size, this model is open-source and promises future optimizations.
On the AI model front, Xiaomi released Mimo V2 Flash, a powerful open-source mixture of experts model with 309 billion parameters optimized for agentic reasoning and coding tasks. Despite its size, only 15 billion parameters are active during use, making it efficient and competitive with top proprietary models like Claude 4.5 and GPT-5. Google also launched Gemini 3 Flash, an efficient multimodal model excelling in image, video, audio understanding, and agentic coding, offering high performance at a fraction of the cost of previous versions. Gemini 3 Flash is now the default in the Gemini app and AI Studio, providing users with a cost-effective yet powerful AI experience.
Finally, several other exciting tools emerged this week, including Lumalabs’ Ray 3 Modify for versatile video reshooting and editing, Cling AI’s motion control for detailed character animation, and Blackforce Labs’ Flux 2 Max image model, which, while good, is overshadowed by OpenAI’s GPT Image 1.5. Additionally, Ego X converts third-person videos into first-person egocentric views, showcasing impressive scene understanding despite some errors. With so many new open-source releases and improvements in AI video, 3D modeling, and animation, this week marks a significant leap forward in accessible, high-quality AI tools for creators and developers alike.
