Chinese artificial intelligence (AI) company DeepSeek has launched Janus Pro 7B, an open-source image generation model, further expanding its footprint in the AI space. The release comes just days after the debut of DeepSeek-R1, the firm’s reasoning-focused language model. The company claims Janus Pro 7B outperforms OpenAI’s DALL-E 3 on multiple benchmarks.
According to a listing on Hugging Face, Janus Pro 7B is an upgrade over DeepSeek’s previous Janus and Janus Pro 1B models. The model leverages an autoregressive framework for multimodal understanding and image generation. It features a split visual encoding system and a unified transformer architecture, aiming for efficiency in processing.
The AI model utilises:
• SigLIP-L vision encoder for image understanding. • Tokeniser with a downsample rate of 16 for improved image generation.
Internal testing by DeepSeek shows Janus Pro 7B scoring 80% on GenEval and 84.2 on DPG-Bench, outperforming models like DALL-E 3 and Stable Diffusion. However, independent testing will determine how it stacks up in real-world applications.
The model is available for download via GitHub and Hugging Face under an MIT license, making it accessible for academic and commercial use. As of now, DeepSeek has not announced an API for the model.
DeepSeek has been making waves with its fully open-source approach, distinguishing itself from competitors like OpenAI and Google. Recently, Perplexity CEO Aravind Srinivas announced that the AI platform now supports DeepSeek-R1, calling it “the world’s most powerful reasoning model.”
Meanwhile, OpenAI CEO Sam Altman acknowledged the competition, stating: “DeepSeek’s R1 is an impressive model, particularly around what they’re able to deliver for the price. We will obviously deliver much better models, and also, it’s legit invigorating to have a new competitor! We will pull up some releases.”
DeepSeek’s rapid advancements have had a notable market impact. On Monday, Nvidia shares dropped by 13%, wiping out approximately $465 billion from its market value - the company’s biggest single-day loss since going public in 1999.
Market analysts speculate that investors are reacting to DeepSeek researchers’ claim that they built R1 without expensive GPUs, at a cost of under $6 million (₹51 crore). This revelation has raised concerns over the future demand for high-end AI chips, which Nvidia dominates.