Fresh off the viral success of its new AI model over the weekend, DeepSeek is already back with another launch this time focusing on AI image generation.The Chinese startup has released an image-generation model called Janus-Pro, aiming to take on US rivals DALL-E 3 and Stable Diffusion. The multimodal model, which generates images from text prompts, is said to outperform competing services in areas such as image quality and accuracy.This launch follows the release of DeepSeek's R1 model, which made waves with its lightning-fast, highly logical responses, and for being trained more quickly and at a fraction of the cost of US models. The model reportedly runs on less advanced Nvidia chips, raising questions about how China is competing without access to cutting-edge US technology. The app recently outpaced ChatGPT in downloads on the Apple App Store.The back-to-back releases signal China's push to gain footing in the growing AI arms race. Meanwhile, last week, President Donald Trump announced a new AI infrastructure initiative, pledging up to $500 million in partnership with OpenAI and other tech firms.The timing also coincides with increased scrutiny of Chinese tech companies, with tensions already high over TikTok's data privacy concerns.Read more:DeepSeek Is the Hot AI App. Don't Get Too Attached to ItJanus-Pro is currently available to download on the AI developer platform Hugging Face.In an introduction on the download page, DeepSeek said: "Janus-Pro surpasses its previous unified model and matches or exceeds the performance of task-specific models. The simplicity, high flexibility, and effectiveness of Janus-Pro make it a strong candidate for next-generation unified multimodal models."The model ranges in size from 1 billion to 7 billion parameters, a key factor in its problem-solving capabilities.The company calls Janus-Pro a "novel autoregressive framework" that solves previous challenges by separating the steps for analyzing and generating images, while still using a single, unified system to process everything."The decoupling not only alleviates the conflict between the visual encoder's roles in understanding and generation but also enhances the framework's flexibility," DeepSeek wrote.