Viral AI company DeepSeek releases new image model family

DeepSeek debuts Janus-Pro models claiming to surpass DALL-E 3.

: DeepSeek releases Janus-Pro, multimodal AI models outperforming DALL-E 3. The largest, Janus-Pro-7B, excels in GenEval and DPG-Bench benchmarks. With sizes from 1 to 7 billion parameters, they offer impressive results despite small image limits. Funded by High-Flyer Capital, DeepSeek's models challenge the U.S. in AI innovation.

DeepSeek, a quickly rising AI company, has introduced a new family of multimodal models named Janus-Pro, purporting superior performance to OpenAI’s DALL-E 3. Available through Hugging Face and licensed under MIT, these models range from 1 to 7 billion parameters, implying improved problem-solving capabilities.

Janus-Pro models employ a novel autoregressive framework for analyzing and creating images, outperforming several other models in benchmarks such as GenEval and DPG-Bench. While most models currently operate on small images up to a resolution of 384 x 384, their compactness alongside impressive performance marks them as leading-edge technology in AI.

DeepSeek's recent prominence also stems from its chatbot's success on the Apple App Store, with funding primarily from High-Flyer Capital Management. Their compute-efficient language models have sparked debates among Wall Street analysts and technologists regarding the future lead in the AI sector and the ongoing demand for AI chips.