Chinese artificial intelligence company DeepSeek unveiled a new open-source image generation model on Monday, stirring interest within the AI community. This follows a series of fully open-source frontier foundation models, including the reasoning-focused DeepSeek-R1. The new model, named Janus Pro 7B, was launched just days after the R1 model and claims to surpass OpenAI’s DALL-E 3 in various benchmarks. Like previous releases, Janus Pro 7B is licensed for both academic and commercial use.
Introduction of DeepSeek Janus Pro 7B AI Model
The Janus Pro 7B has been detailed in a listing on Hugging Face page. This model succeeds the earlier Janus and Janus Pro 1B models, featuring significant upgrades to its functionality. DeepSeek describes it as an autoregressive framework that integrates multimodal understanding and generation, alongside various improvements made to the model’s architecture and encoder.
Prioritizing efficiency, the Janus Pro 7B separates visual encoding into distinct pathways while maintaining a unified transformer architecture for processing tasks. The model employs the SigLIP-L vision encoder for multimodal understanding and incorporates a tokeniser that operates with a downsample rate of 16 for generation purposes.
According to internal testing released by DeepSeek, Janus Pro 7B achieved scores of 80 percent on the GenEval benchmark and 84.2 on DPG-Bench. Both DALL-E 3 and Stable Diffusion have reported lower scores in these assessments. Independent evaluations in the near future are expected to provide further clarity on Janus Pro 7B’s performance.
To download the model, users can access it from GitHub here as well as on Hugging Face, where it is available under an MIT license. A demonstration of the AI model can also be accessed here. Currently, DeepSeek has not introduced an application programming interface (API) for this model.
Perplexity Expands Support for DeepSeek-R1
On the same day, Aravind Srinivas, CEO of Perplexity, announced that the AI platform would now accommodate DeepSeek-R1 in addition to OpenAI’s o1 AI model. He referred to DeepSeek-R1 as the “world’s most powerful reasoning model,” affirming that it will be accessible to all users.
Though there are currently restrictions on the number of outputs generated using the model, the company has intentions to increase this limit. They also emphasized that the model is hosted in the United States to mitigate concerns over data being transmitted to Chinese servers.
In a separate development, OpenAI CEO Sam Altman finally addressed the recent rise of DeepSeek’s AI models, calling the R1 model “impressive” given the value it offers. He pointed out that o1’s API prices are significantly higher compared to those of R1.
“We will obviously deliver much better models and also it’s legit invigorating to have a new competitor! We will pull up some releases,” Altman further stated.
The same day, shares of Nvidia dropped approximately 13 percent, resulting in a loss of about $465 billion (around Rs. 40 lakh crore) from the company’s market capitalization. This marks the largest single-day decline for the tech giant since its public listing in 1999.
Market analysts have suggested that the decline might stem from investor concerns regarding DeepSeek’s claims. The researchers stated in a paper that they were able to develop the R1 model without the reliance on expensive GPUs for a total cost of under $6 million (approximately Rs. 51 crore).
