The AI services being launched or announced continue to surprise us every week. We’ve compiled the most notable ones of the week for you.
HUNYUAN3D 3.0
Tencent released Hunyuan3D 3.0, offering 3 times higher precision and 1536° ultra HD voxel modeling. It can capture missing details, realistic facial features, and professional-grade textures for gaming, film, and e-commerce applications. https://hunyuan-3d.com/
WAN2.2
Wan introduced Wan2.2, a 5B parameter video diffusion model with an MoE architecture that offers higher capacity for the same cost. It can deliver cinema-quality visuals, complex motion generation, and efficient 720p text-to-video and image-to-video output at 24 fps. https://wan.video/
MOONDREAM 3
Moondream 3 launched as a 9B parameter, 2B active MoE visual-language model, providing state-of-the-art visual reasoning in a compact, application-friendly design. https://moondream.ai/
SRPO
Tencent-Hunyuan unveiled SRPO, a diffusion fine-tuning method that stabilizes the training process, corrects noisy images, and shortens computation time. This method enables faster optimization, prevents reward hacking, and supports controllable style adjustments for models like FLUX.1.dev. https://www.srpo.net/
REVE IMAGE
Reve launched Reve Image, which combines image generation, restyling, a drag-and-drop editor, a creative assistant, and a beta API. Users can create and edit images with natural language and integrate Reve’s capabilities into their own applications. https://app.reve.com/
LING-FLASH 2.0
Ling-flash-2.0 is now open-source and is a 100B parameter MoE LLM with 6.1 billion active parameters. Trained on 20T+ tokens, it exhibits near-perfect performance in complex reasoning, code generation, and frontend development, making it the most advanced among dense models under 40 billion parameters. https://huggingface.co/inclusionAI/Ling-flash-2.0
VOXCPM
VoxCPM, a tokenizer-free TTS model powered by MiniCPM-4, offers zero-shot voice cloning and hyper-realistic speech with natural harmony. Trained with over 1.8 million hours of data, it achieves state-of-the-art performance. https://voxcpm.com/
UMO
UMO, a unified multi-identity optimization framework for image customization, was introduced. It can ensure high identity consistency, reduce entanglement among multiple reference images, and will be fully open-source with models, scripts, and training code. https://bytedance.github.io/UMO/
RAY3
Luma AI introduced Ray3, the first reasoning video model capable of producing studio-quality HDR. Its new draft mode enables fast iteration with improved physics and coherence and is now available for free in Dream Machine. https://lumalabs.ai/ray
PAPER2AGENT
The newly announced Paper2Agent infrastructure automatically converts academic papers into active AI agents. Using multiple sub-agents, the system builds a robust Model Context Protocol (MCP) from a paper’s text and code, enabling the resulting agent to apply the paper’s methods and data to new projects. https://github.com/jmiao24/Paper2Agent
You Might Also Like;
- We Selected 10 Series Similar to Stranger Things for Those Who Love It
- Where and How is Silver Used in Electric Vehicles?
- Hyundai Unveils Its Multi-Purpose Wheeled Robot
Follow us on TWITTER (X) and be instantly informed about the latest developments…
