In the early hours of January 28th, the highly anticipated AI community Hugging Face revealed a groundbreaking announcement: DeepSeek has launched an open-source multimodal AI model called Janus-Pro. This model is available in two parameter sizes, 1 billion and 7 billion, and was trained using only 128 NVIDIA A100 GPUs over the course of a week. Notably, in benchmark tests such as GenEval and DPG-Bench, Janus-Pro-7B demonstrated outstanding performance, surpassing OpenAI’s DALL-E 3 and Stable Diffusion models.
In short, the Janus-Pro model integrates multiple functionalities, enabling AI to not only interpret images (based on SigLIP-L technology) but also generate images (drawing from LlamaGen). Additionally, the model is available in two sizes, 1.5B and 7B, to meet different needs. It’s worth mentioning that although GPT-4o has garnered attention in the multimodal image generation field, its related models have yet to be released publicly. In contrast, Janus-Pro undoubtedly brings new possibilities and choices to the industry.
What is DeepSeek
DeepSeek, officially known as Hangzhou DeepSeek Artificial Intelligence Fundamental Technology Research Co., Ltd., was founded on July 17, 2023. It is an enterprise dedicated to technological innovation, focusing on the development of cutting-edge large language models (LLM) and related technologies, with the goal of achieving breakthroughs in the field of artificial intelligence.
At the end of 2024, DeepSeek proudly launched its new generation large language model, V3, and announced that it would be open-sourced for global developers to collaborate on research and progress. After rigorous testing, the V3 model achieved excellent results in several benchmarks, even surpassing some mainstream open-source models. What’s even more noteworthy is that, alongside its outstanding performance, the V3 model also offers significant cost advantages, undoubtedly laying a solid foundation for its widespread application in the market.
DeepSeek’s latest model
Let’s dive deeper into DeepSeek’s latest model, which is actually an advanced version and successor of Janus and JanusFlow.
More specifically, this model is built upon DeepSeek-LLM-1.5b-base/DeepSeek-LLM-7b-base and is a multimodal large-scale model that integrates both understanding and generation capabilities. The entire model uses an autoregressive framework, and its innovation lies in overcoming the limitations of previous methods by decomposing visual encoding into separate paths. At the same time, it maintains the use of a single, integrated transformer architecture for task execution.
This decomposition not only effectively resolves the role conflict between the visual encoder’s functions of understanding and generation, but also significantly enhances the flexibility and adaptability of the entire framework.
Compared to the previous version, Janus, DeepSeek’s new model shows significant improvements in performance. For short prompts, it provides more stable outputs, meaning the model’s responses are more reliable and consistent when processing user inputs.
Additionally, the new model exhibits higher visual quality, with generated images or videos appearing clearer and more detailed. Not only that, it also offers richer detail representation, capturing more subtle nuances and making the generated content more vivid and lifelike.
It’s worth noting that this new model also has the ability to generate simple text, a feature that was not available in previous versions, undoubtedly enhancing its practicality and flexibility.
DeepSeek’s Cost Advantage: Surpassing Trump’s $50 Million AI Blueprint
Among the many models that claim to rival the GPT series, why has DeepSeek managed to create such a significant stir in Silicon Valley’s AI scene?
The key lies in the astonishingly low cost of its models. Take DeepSeek-V3, for example. This model, comparable to GPT-4, uses only 2,000 Nvidia chips, with a total training cost of less than $5.58 million—just one-tenth of the cost of similar models.
In comparison, OpenAI’s ambitious “interstellar blueprint,” as described by Trump, cost a staggering $50 million. The cost-effectiveness of DeepSeek is therefore self-evident.
President Trump: “The release of DeepSeek AI from a Chinese company should be a wake-up call for our industries that we need to be laser focused on competing to win.”
Conclusion
In conclusion, DeepSeek has emerged as a prominent player in the AI industry, impressively launching the Janus-Pro model, an open-source multimodal AI that integrates image interpretation and generation capabilities. The model’s benchmark test results surpass leading models like DALL-E 3 and Stable Diffusion, showcasing DeepSeek’s technological prowess.
Moreover, DeepSeek’s cost-effective approach to AI development, particularly evident in the V3 model, positions it as a disruptive force in the market. By offering advanced AI solutions at a fraction of the cost of its competitors, DeepSeek is not only reshaping the AI landscape but also setting a new benchmark for cost-efficiency in the industry.
Leave a Comment