Less than a week after sending tremors through Silicon Valley and across the media landscape with an affordable large language model called DeepSeek-R1, the Chinese AI startup behind that technology has debuted another new product — the multimodal Janus-Pro-7B with an aptitude for image generation. Further mining the vein of efficiency that made R1 impressive to many, Janus-Pro-7B utilizes “a single, unified transformer architecture for processing.” Emphasizing “simplicity, high flexibility and effectiveness,” DeepSeek says Janus Pro is positioned to be a frontrunner among next-generation unified multimodal models.

According to DeepSeek, Janus-Pro “matches or exceeds the performance of task-specific models,” the company said on Hugging Face, where the models and a demo are hosted. The code is also available on GitHub.

“Janus-Pro is a unified understanding and generation MLLM, which decouples visual encoding for multimodal understanding and generation,” explains DeepSeek. That summary is in contrast to the near hysteria triggered by R1’s release last week. VentureBeat notes that event prompted an “AI stock bloodbath, igniting fresh fears of Chinese tech dominance.”

This unexpected second act from DeepSeek “intensifies investor worries about China’s growing power in AI and further pressures American tech companies,” VentureBeat writes, noting Janus Pro’s debut precipitated a fresh selloff of U.S. AI stocks, “timing that appears to be deliberate and designed to highlight the Beijing-based firm’s challenge to Silicon Valley.

Recent events suggest Chinamania is gripping the U.S., where the fate of TikTok has preoccupied the Washington elite even as the rank and file pass over American competitors like Instagram, X and YouTube to make a virtually unknown Chinese social app called RedNote “the top downloaded app in the U.S.” according to the Associated Press.

With Janus-Pro, DeepSeek challenges U.S. rivals like Midjourney, Stable Diffusion and OpenAI’s DALL-E 3. CNET reports Janus “is said to outperform competing services in areas such as image quality and accuracy.”

Janus-Pro-7B is available in one-billion and seven-billion parameter models, reports Tom’s Guide, which says that “according to DeepSeek, the latter is able to compete with both Stable Diffusion and DALL-E 3 in benchmarking tests,” though this has yet to be rigorously tested by third parties.

Like R1, “Janus-Pro is under an MIT license, meaning it can be used commercially without restriction,” explains TechCrunch.

Related:
DeepSeek Scrambles U.S.-China Tech War, The Wall Street Journal, 1/29/25
DeepSeek’s AI Avoids Answering 85% of Prompts on ‘Sensitive Topics’ Related to China, TechCrunch, 1/29/25
DeepSeek’s Big Question: Where Does AI’s True Value Reside?, The Wall Street Journal, 1/29/25
It’s Not Just DeepSeek. A Guide to the Chinese AI Companies You Need to Know, The Wall Street Journal, 1/29/25