Is DeepSeek’s new image model another win for cheaper AI?

Chinese AI startup DeepSeek isn’t squandering its momentum anytime soon. 

Just moments after knocking ChatGPT out of the top spot in the App Store for most downloaded free apps, the company released Janus-Pro’s multimodal text-to-image AI model on Monday. Like R1, DeepSeek’s flagship model, Janus-Pro is open source under an MIT license (making it commercially viable) and downloadable via HuggingFace and GitHub. 

Also: I tested DeepSeek’s R1 and V3 coding skills – and we’re not all doomed (yet)

Similar to the R1 release, DeepSeek launched several versions of Janus-Pro, ranging from 1B to 7B-parameters in size. DeepSeek’s own testing claims that Janus-Pro-7B, the larger of the two, beats established image generators like Stable Diffusion and DALL-E on the GenEval and DPG-Bench benchmarks. 

DeepSeek says that the model uses an “autoregressive framework” and “surpasses” unified models. 

Janus-Pro builds on Janus, its original version released last year, and can create and analyze images. Smaller-parameter models in the family are limited to analyzing images of 384 x 384 resolution, which is a drawback. 

That said, Janus-Pro’s performance is still competitive, especially given DeepSeek’s reportedly lower training costs when compared with those of US-based AI companies. In December, a company research paper claimed its V3 model only cost $5.6 million to make, which would be fraction of what Google and OpenAI have spent on their star models. Some have expressed concern that this number is incomplete (leaving out R&D, data, and personnel costs) or hard to believe.

Nvidia even told CNBC that the model is “an excellent AI advancement.” In the context of DeepSeek’s other rapid-fire releases, the model family’s first impressions are mixed but overall positive. These may shift as more users test Janus-Pro for themselves against other image models. 

Also: Apple researchers reveal the secret sauce behind DeepSeek AI

ZDNET is also looking into reports that DeepSeek’s approach is more energy efficient than its US counterparts, which would be another significant shakeup for the AI industry and investment in the space. The release of Janus-Pro calls into question plans like Stargate, a $500 billion initiative between several AI giants and touted by the Trump administration, given that competitive AI may not require the energy and scale of the initiative’s proposed data centers.