The rapidly rising Chinese AI startup DeepSeek has unveiled a suite of new multimodal AI models, which the company claims outperform OpenAI’s DALL-E 3.
The neural networks are part of the new Janus-Pro family, with sizes ranging from 1 billion to 7 billion parameters.
Janus-Pro is capable of analyzing and generating new images. In two benchmarks, GenEval and DPG-Bench, it surpassed DALL-E 3, PixArt-alpha, Emu3-Gen, and other counterparts.
The models demonstrate high performance despite their relatively small size.
“Janus-Pro surpasses previous unified models and is comparable to or exceeds neural networks specializing in individual tasks. The simplicity, high flexibility, and efficiency of Janus-Pro make it a strong candidate for creating the next generation of unified multimodal neural networks,” stated DeepSeek.
The startup previously made headlines with the release of the R1 model, which led to a decline in the shares of Nvidia, several other tech companies, and cryptocurrencies.
Experts have urged caution in drawing conclusions about the real capabilities of the neural network.