DeepSeek, a Chinese artificial intelligence (AI) firm, released an open-source image generation model on Monday. The company has generated a buzz in the AI space recently by releasing multiple fully open-source frontier foundation models, including the reasoning-focused DeepSeek-R1. The new model, dubbed Janus Pro 7B, comes days after the R1 model and the company claims it outperforms OpenAI’s DALL-E 3 in multiple benchmarks. Similar to its previous releases, the new AI model is also available with a permissive licence for both academic and commercial usage.
Table of Contents
DeepSeek Janus Pro 7B AI Model
The new AI model has been detailed in a Hugging Face listing. The model is the successor to the Janus and Janus Pro 1B models, and the company has made significant upgrades to its functionality. Describing it as an autoregressive framework that unifies multimodal understanding and generation, DeepSeek stated that several changes were made to the architecture as well as the encoder to improve the model’s capabilities.
With a focus on efficiency, the DeepSeek Janus Pro 7B now decouples visual encoding into separate pathways while using a unified transformer architecture for processing. For multimodal understanding, it uses the SigLIP-L vision encoder, while for generation, it uses a tokeniser with a downsample rate of 16.
Based on internal testing shared by the company, Janus Pro 7B scored 80 percent on GenEval and 84.2 on DPG-Bench benchmarks. Both DALL-E 3 and Stable Diffusion models have scored less on these tests. However, independent testing in the coming days will reveal a better picture of Janus Pro 7B’s capabilities.
Currently, it is available to be downloaded from both GitHub’s listing page as well as on Hugging Face. It is available with an MIT licence. A demo of the AI model can be found here. DeepSeek has not announced any application programming interface (API) for the model so far.
Perplexity Adds Support for DeepSeek-R1
Aravind Srinivas, the CEO of Perplexity, announced on Monday that the AI platform will now support DeepSeek-R1 alongside OpenAI’s o1 AI model. Calling it the “world’s most powerful reasoning model,” he said the model will be available to all users.
Currently, there is a limit to the number of outputs that can be generated using the model, but the company plans to increase it. Notably, the company highlighted that the model is being hosted in the US, to alleviate any concerns regarding the data being sent to Chinese servers.
Separately, OpenAI CEO Sam Altman finally commented on the sudden rise of DeepSeek’s AI models. He called the R1 model “impressive”, especially for what the company is “able to deliver for the price.” Notably, o1’s API prices are set much higher than R1’s.
“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 added.
On Monday, Nvidia shares dropped by around 13 percent, wiping out around $465 billion (roughly Rs. 40 lakh crore) from the company’s market capitalisation. This is the biggest single-day drop for the tech giant ever since it went public in 1999.
Market experts speculated that the fall could be because investors are concerned about the claims of DeepSeek researchers. The researchers, in a paper, had said that they were able to build R1 without expensive GPUs and at a cost of under $6 million (roughly Rs. 51 crore).
