For two years, OpenAI was seen as a leading company in the AI sector. Microsoft secured its technology through a billion-dollar deal, and the company, led by CEO Sam Altman, is part of Donald Trump’s Stargate project, which could cost up to 500 billion US dollars. However, almost overnight, a challenger from China named Deepseek seems to have surpassed not only OpenAI but also other companies like Meta, causing concern for these manufacturers.
Deepseek’s Janus-Pro-7B model is claimed to be better than Dall-E and Stable Diffusion. Initially, the Deepseek R1 model surpassed OpenAI’s counterparts not only in benchmarks but also pushed ChatGPT to second place in app store charts. Now, Deepseek introduces the Janus-Pro family of multimodal models, claiming they outperform the competition. According to Deepseek on Hugging Face, “Janus-Pro surpasses previous unified models and matches or exceeds the performance of task-specific models.”
According to TechCrunch and Reuters, Deepseek claims that its largest model, Janus-Pro-7B, has outperformed OpenAI’s Dall-E3 and Stability AI’s Stable Diffusion in two AI benchmarks for image generation from text prompts.
The Janus-Pro models use between one and seven billion parameters. For comparison, OpenAI’s Dall-E uses twelve billion parameters. Deepseek’s models can not only create images from prompts but also analyze them. The company demonstrates that more parameters are not always better with its chatbot, which theoretically has 671 billion parameters available but only uses 37 billion in practice, requiring significantly fewer hardware resources. TechCrunch praises the performance of Janus-Pro relative to its model size as “impressive.”
However, the image analysis only works with material that has a resolution of 384 x 384 pixels or less. The predecessor model, simply called Janus, is said to be surpassed by the Pro variants through improvements in training processes, data quality, and model size, leading to enhanced image stability and more detail, according to Reuters.
How the image AI will perform in practice remains to be seen. In a brief test, the Deepseek chatbot appeared to be aligned with the “social morals of China.” As a result, the chatbot refused to provide an answer to a question about the Tiananmen Square incident. It is possible that the image AI is subject to similar restrictions.
The Janus-Pro models are available under the MIT license, allowing them to be used commercially without restrictions. They can be downloaded for free from Hugging Face and GitHub.
Overall, Deepseek’s rise as a challenger in the AI field highlights the rapid advancements and competition in this technology sector. While OpenAI has been a dominant force, new players like Deepseek are showing that innovation can come from unexpected places, potentially reshaping the landscape of AI development.