After causing a true cataclysm in technological ones dedicated mainly to the manufacture of chips for AI, with its V3 and R1 models, DeepSeek has returned to the load with a New family of multimodal models. He has called them Janus-Pro and ensures that their results exceed those who can achieve various multimodal models already in use, with a markedly lower size than them.
Available to download through the Hugging Face development platform, they count, depending on their dimensions, with between 1,000 and 7,000 million parameters. Like the code of the other models of the company, that of the Janus-Pro family is also Open Source (MIT license), which means that they can be modified and adapted to improve them, even for commercial purposes, without No type of restriction.
The Models of the Janus-Pro family are, according to Deepseek herself, a new self-representive framework, capable of analyzing images and creating new ones. The most powerful family model, the Janus-Pro-7B, exceeds the results obtained by models such as Openai Dall-E3, Stable Diffusion XL of Stability AI, EMU3-GNE or Pixart-Alpha. Of course, the smallest models of the family, those that have the lowest number of parameters, are only able to analyze small images, with a maximum resolution of 384 × 384.
The company has confirmed the launch through a post on the Hugging Face website, in which they emphasize that Janus-Pro exceeds the unified Price, and covers or exceeds Deepseek’s expectations in terms of the performance of models that are responsible for perform concrete tasks.
In addition, they point out that «The simplicity, high flexibility and effectiveness of Janus-Pro makes its models candidates for the next generation of next-generation unified multimodal models«.
Deepseek is therefore determined to break the schemes of the companies dedicated to AI in the US and Europe, which after seeing how hardware manufacturers for AI, and also several semiconductors, such as Broadcom, TSMC or NVIDIA; They lost billions in the stock market in yesterday’s session, begin to react. Openai, for example, has already pointed out that they will launch more and better models, and points towards an acceleration in its road map to compete with the Chinese startup.
In Nvidia, despite losing about 600,000 million dollars yesterdayapparently they keep calm, and qualify Deepseek-R1 as «An excellent advance in AI and a perfect example of time scaling test, whose work illustrates how new models can be created using such technique, taking advantage of available models and computing that perfectly meets export controls«. We will have to see how the bags evolve, and the sector, in the next few days.
