DeepSeek’s Janus-Pro Challenges DALL-E 3 with Compact AI Power

DeepSeek has launched a new set of multimodal AI models called Janus-Pro, which it claims are better than OpenAI’s DALL-E 3.

Key Points

• DeepSeek multimodal AI models, are available for download from the AI dev platform Hugging Face, ranging in size from 1 billion to 7 billion parameters.
• Janus-Pro is under an MIT license, meaning it can be used commercially without restriction.
• DeepSeek’s Janus-Pro, described as a “novel autoregressive framework,” can analyze and create images.
• According to the company, Janus-Pro-7B outperforms DALL-E 3 and others like PixArt-alpha and Stable Diffusion XL on two AI benchmarks, GenEval and DPG-Bench.
• Despite handling smaller images (384×384), Janus-Pro compact size, flexibility, and strong performance make it a promising next-gen multimodal model.

Background

Recenctly Deepseek launched Its AI reasoning model R1, to improve problem-solving and analytical skills.

The company also introduced six smaller open-source models based on R1, with two performing similarly to OpenAI’s OpenAI-01 mini.

DeepSeek  gained significant attention this week after its chatbot app overtaken rival ChatGPT to become the top-rated free application available on Apple’s App Store in the United States.

DeepSeek’s language models have prompted many Street analysts and technologists to question whether the U.S. can maintain its lead in the race and whether the demand for AI chips will continue.

News Gist

DeepSeek’s new Janus-Pro multimodal AI models, available on Hugging Face, range from 1 to 7 billion parameters and are MIT-licensed for unrestricted commercial use.

Janus-Pro-7B outperforms DALL-E 3 and other AI models on benchmarks like GenEval and DPG-Bench.

Despite handling smaller images (384×384), its compact size, flexibility, and strong performance position it as a next-gen multimodal model.