DeepSeek, a Chinese artificial intelligence (AI) company, released an open source image generation model on Monday. The company recently caused a sensation in the AI space by releasing a variety of fully open source border basic models including DeepSeek-R1, which focuses on reasoning. The new model, named as the Janus Pro 7b, is a few days after the R1 model and the company claims to outperform Openai’s Dall-E 3 in multiple benchmarks. Similar to previous releases, the new AI model also has permissioned licenses for academic and commercial usage.
DeepSeek Janus Pro 7B AI Model
The new AI model has been detailed in the hug’s face list. The model is the successor to the Janus and Janus Pro 1B models, and the company has made significant upgrades to its capabilities. DeepSeek describes it as an automated regression framework that unified multimodal understanding and produces, and represents some changes to the architecture and encoder to improve the functionality of the model.
With efficiency-focused, DeepSeek Janus Pro 7b now breaks down visual coding into separate pathways while processing using a unified transformer architecture. For multimodal understanding, it uses a siglip-l visual encoder, and for generation, it uses a token with a sample rate of 16.
According to internal tests shared by the company, the Janus Pro 7B scored 80% on Geneval, while the DPG base benchmark was 84.2. In these tests, DALL-E 3 and stable diffusion models were scored less. However, independent testing in the coming days will reveal better situations of the Janus Pro 7b’s capabilities.
Currently, it can be downloaded from Github’s list page as well as from the hugged face. It can come with a MIT license. A demonstration of the AI model can be found here. To date, DeepSeek has not announced any application programming interfaces (APIs) for this model.
Confusion adds support for DeepSeek-R1
Confused CEO Aravind Srinivas announced on Monday that the AI platform will now support DeepSeek-R1 along with OpenAI’s O1 AI model. He called it “the most powerful inference model in the world”, which he said will work for all users.
Currently, there are limits on the number of outputs that can be generated using the model, but the company plans to increase it. It is worth noting that the company stressed that the model is being hosted in the United States to alleviate any concerns about data sent to Chinese servers.
Additionally, Openai CEO Sam Altman finally commented on the sudden rise of the DeepSeek AI model. He called the R1 model “impressive”, especially because the company’s prices are able to be delivered at prices. “It is worth noting that the API price of O1 is much higher than R1.
“Obviously, we will provide better models and have legitimate people who can have a new competitor! We will extract some distributions,” Altman added.
NVIDIA shares fell about 13% on Monday, eliminating about US$465 billion (about Rs 4 million) from the company’s market capitalization. This is the biggest single-day decline for tech giants since they went public in 1999.
Market experts speculate that the fall may be because investors care about the claims of DeepSeek researchers. The researchers said in a paper that they were able to build the R1 without an expensive GPU cost less than $6 million (about Rs 5.1 billion).
