This story incorporates reporting from decrypt, Business Insider, gadgets360 and The Gazette.
DeepSeek, a China-based artificial intelligence lab, has once again raised industry standards with the release of its latest image generation model, Janus Pro. The company claims that this new model surpasses leading names like OpenAI’s DALL-E 3 in multiple benchmarks. Released as open-source with a permissive license, Janus Pro allows for both academic and commercial use worldwide. The announcement, made on a recent Monday, builds on DeepSeek’s momentum from previous breakthroughs, including their R1 model which was noted for its advanced reasoning capabilities.
The Janus Pro model employs a novel autoregressive framework, which decouples visual encoding into separate pathways yet maintains a unified transformer architecture. This design, according to DeepSeek’s documentation, marks a significant leap forward in image generation technology. DeepSeek asserts that Janus Pro not only matches but often exceeds the performance of task-specific models while maintaining a singular, coherent structure. This positions Janus Pro as a formidable contender for future developments in unified multimodal models.
Internal testing by DeepSeek reported that Janus Pro 7B scored 80 percent on the GenEval benchmark and 84.2 on the DPG-Bench benchmarks, which are industry-recognized tests for evaluating image generation models. These figures demonstrate notable advancements in both image stability and detail richness compared to prior models. DeepSeek’s technical report highlighted these improvements, attributing the success to enhancements in training processes, data quality, and model size.
Furthermore, the model’s capabilities have been benchmarked against base, non fine-tuned models from competitors like Stability AI’s Stable Diffusion and Pixart Alpha. DeepSeek insists these comparisons favor Janus Pro, showcasing its capacity to outperform under standard conditions without additional model tuning.
The introduction of Janus Pro has generated significant attention in the tech sector, influencing market dynamics. Following the announcement, key tech stocks, including Nvidia and Oracle, experienced notable declines. This reflects growing investor awareness of the competitive disruption posed by DeepSeek’s innovations in the AI landscape.
DeepSeek’s consistent release of high-performance models, paired with its decision to open-source these innovations, underscores its strategy to democratize access to cutting-edge AI technology. This move not only cements DeepSeek’s position as a disruptor in the field but also aligns with broader trends favoring open-source collaboration in AI development. Conversely, OpenAI and Stability AI have not yet commented on DeepSeek’s claims, leaving the industry’s response to these new benchmarks to be seen.
Quartz Intelligence Newsroom uses generative artificial intelligence to report on business trends. This is the first phase of an experimental new version of reporting. While we strive for accuracy and timeliness, due to the experimental nature of this technology we cannot guarantee that we’ll always be successful in that regard. If you see errors in this article, please let us know at qi@qz.com.