• Source:JND
Add as a preferred

Janus Pro 7B: DeepSeek, a Chinese AI startup, made waves in the tech world with the release of its open-source image generation model, Janus Pro 7B, on Monday. This comes shortly after the launch of its DeepSeek-R1 reasoning model, both of which are designed to push the boundaries of artificial intelligence. The company has been gaining attention for its commitment to releasing fully open-source models, including those with advanced capabilities, available for both academic and commercial use.

Janus Pro 7B, a successor to the earlier Janus and Janus Pro 1B models, is being touted for its improvements in performance. It features an autoregressive framework that merges multimodal understanding with generation, making it a versatile tool in AI development. The architecture and encoder have been significantly enhanced to boost the model's efficiency. The Janus Pro 7B uses a unique approach for visual encoding, which is now separated into distinct pathways. Its unified transformer architecture processes data effectively, incorporating the SigLIP-L vision encoder for multimodal understanding and a tokeniser with a downsample rate of 16 for content generation.

ALSO READ: iPhone 17 Pro Max: What Can We Expect In Terms Of Price, Design, And Camera

 

In internal tests, the Janus Pro 7B achieved scores of 80 percent on GenEval and 84.2 on DPG-Bench, outperforming OpenAI’s DALL-E 3 and Stable Diffusion in these benchmarks. While these results are promising, further independent testing is expected to provide a more comprehensive evaluation of the model's capabilities.

Currently, DeepSeek has made Janus Pro 7B accessible for download via GitHub and Hugging Face, offering it with an MIT license. The model is available for testing but does not yet include an API. The company has also not yet revealed any plans for an API for broader integration.

Meanwhile, DeepSeek’s DeepSeek-R1 reasoning model has earned additional attention. Perplexity, an AI platform, has announced that it will support this model alongside OpenAI’s own o1 AI model. Aravind Srinivas, CEO of Perplexity, referred to DeepSeek-R1 as “the world’s most powerful reasoning model,” and confirmed it will be made available to all users.

Despite some limitations in output generation, DeepSeek plans to increase the number of outputs in the future. The company has also made it clear that the model is hosted in the US to address any concerns about data being stored on Chinese servers.

The Janus Pro 7B and DeepSeek-R1 models are a testament to DeepSeek's rapidly advancing AI capabilities, signaling its growing presence in the global AI landscape.

ALSO READ: China's Startup DeepSeek AI: Rapid Success Followed By 'Malicious Attacks'

 


Also In News