Chinese AI Firm DeepSeek Challenges OpenAI with Open-Source Image Generation Model
DeepSeek, a rising star in the Chinese artificial intelligence (AI) landscape, has intensified its challenge to industry leader OpenAI. They have released Janus Pro 7B, an open-source image generation model. The company claims it outperforms OpenAI’s DALL-E 3 in several benchmarks. This move follows the recent release by DeepSeek of the fully open-source, reasoning-focused DeepSeek-R1 model. It signals the company’s aggressive push in the AI arena.
Janus Pro 7B follows DeepSeek’s previous Janus models. The company describes it as an autoregressive framework. It unifies multimodal understanding and generation. DeepSeek has implemented significant architectural changes and encoder improvements to boost the model’s capabilities. A key focus has been efficiency, with visual encoding decoupled into separate pathways while a unified transformer architecture handles processing. The model uses the SigLIP-L vision encoder to understand multiple modalities. It employs a tokenizer with a downsample rate of 16 for generation.
DeepSeek’s internal testing suggests Janus Pro 7B achieves an 80% score on GenEval and 84.2 on DPG-Bench, surpassing the scores of both DALL-E 3 and Stable Diffusion on these benchmarks. However, independent evaluation is crucial to validate these claims and provide a more comprehensive assessment of the model’s performance.
Janus Pro 7B, like DeepSeek’s other recent releases, is available under a permissive MIT license. This license allows both academic and commercial use. The model can be downloaded from GitHub and Hugging Face, and a demo is available online. DeepSeek has not yet announced plans for an API.
DeepSeek’s Momentum Builds with Perplexity Support and Altman’s Acknowledgment
DeepSeek’s momentum is further amplified by Perplexity’s announcement that it now supports DeepSeek-R1 alongside OpenAI’s o1 model. Perplexity CEO Aravind Srinivas hailed R1 as “the world’s most powerful reasoning model.” He made it available to all users on the platform. While output generation is currently limited, Perplexity plans to increase this capacity. The company emphasized that the model is hosted in the US, addressing potential data security concerns.
Adding to the buzz, OpenAI CEO Sam Altman has finally commented on DeepSeek’s recent surge. He acknowledged the R1 model as “impressive,” particularly praising its price-to-performance ratio, a clear jab at o1’s higher API pricing. Altman added, “We will obviously deliver much better models and also it’s legit invigorating to have a new competitor! We will pull up some releases.” This acknowledgment from Altman underscores the growing recognition of DeepSeek’s potential to disrupt the established AI landscape. The company’s commitment to open-source models is noteworthy. Its focus on image generation and reasoning capabilities strengthens its position. DeepSeek is a formidable challenger to OpenAI and other leading AI firms.
Leave a Reply