While Wall Street’s tech stocks experienced a dramatic plunge on January 28, a new star in China’s AI sector was illuminating the entire industry with its disruptive brilliance—the DeepSeek team’s officially open-sourced Janus-Pro series model not only redefined the performance boundaries of multimodal large models but also showcased China’s hardcore strength in AI to the world with textbook-level architectural innovation.
1. On the Night of the Stock Market Meltdown, a Chinese Team is “Stir-Frying AI” in the Kitchen
As the tech sector of the US stock market collectively plummeted, Janus-Pro-7B crushed MetaMorph with an MMBench score of 79.2, outscoring DALL-E 3 (0.67) and SD3 (0.74) with a GenEval score of 0.80. This technical brilliance blooming amidst a capital market winter confirms that the DeepSeek team’s foundational breakthroughs are far from a fleeting marketing gimmick.
The core breakthrough lies in the “visual encoding decoupling architecture”—this seemingly simple design decision effectively addresses the “Theseus’s Ship Paradox” that has plagued the industry for years:
-
Traditional models attempt to use the same visual encoder for both understanding and generation, akin to asking the same brain to be both a rigorous mathematician and a free-spirited artist.
-
Janus-Pro employs a dual-path design of SigLIP-L semantic encoder + VQ discretization generator, enabling the model to transform into a logically rigorous “scientist” for understanding tasks and a wildly imaginative “artist” for generation tasks.
This paradigm shift at the architectural level allows the 7B parameter Janus-Pro to achieve nearly a 10-point improvement in MMBench understanding tasks compared to its predecessor, while also comprehensively surpassing Stable Diffusion in image generation quality. While other vendors are still competing on data volume and computational power, DeepSeek has already ascended to a higher dimension of architectural innovation.
2. Training Strategy: The AI Industry’s “Old Ganma Recipe” Revealed
If architectural innovation is the skeleton, then Janus-Pro’s three-stage training strategy is the genetic engineering that gives it a soul:
-
ImageNet Foundation Stage (200 epochs of extended training): Focused on pixel dependency modeling with fixed LLM parameters, honing basic strokes like a top-tier artist.
-
Real Data Transition Stage: Decisively abandoning synthetic data reliance, reshaping the generation logic with 72M real aesthetic data, achieving photographic-level detail in images.
-
Dynamic Balance Stage (5:1:4 data ratio): Finding the golden ratio among multimodal understanding, pure text dialogue, and image generation, balancing logical rigor with creative freedom.
The revolutionary effects of this training strategy are astonishing: the 7B model improved its short prompt response speed by 300% compared to the 1.5B version, and the PSNR metric for generated images increased by 15.8%, while training efficiency was optimized by 40%. This breakthrough of “wanting it all” is the best testament to DeepSeek’s engineering prowess.
3. The Chinese Team’s Mysterious Buff: Instant Noodles + Grind = Black Technology
In a Shenzhen laboratory at 4 AM, programmers slurp red-braised beef instant noodles while feeding the AI 200 panda emoji images—this seemingly absurd “Wednesday Confusion Behavior” led the model to suddenly illustrate Einstein’s theory of relativity with panda emojis by Friday.
Looking at the list of paper authors—Chen Xiaokang, Wu Zhiyu, Liu Xingchao… all with Pinyin names, this plays out like a real-life version of “The Wandering Earth”:Even more exciting is that the list of authors is entirely composed of Chinese research teams, proving that China does not lack top AI talent. When researchers like Chen Xiaokang and Wu Zhiyu submit code on GitHub, they are not just tapping on keyboards but also beating the drum for China’s ascent to the pinnacle of AI.
Supporting this cyber comedy is substantial technological progress: the model parameters surged from 1.5B to 7B, with a 300% increase in generation speed, and astonishingly, the error rate directly surpassed Stable Diffusion by 62%. While peers are still competing on computational power, this group has already marked a magical realism chapter in AI evolution with instant noodles and emoji images.
4. Open Source Strike: Rubbing Technology Hegemony into the Ground
In an industry context where OpenAI is gradually closing off and Stability AI is mired in commercialization, Janus-Pro’s choice to go fully open source is nothing short of a revolution:
-
Completely Open Model Architecture: From SigLIP-L encoder to VQ generator, all modules can be freely disassembled and reassembled.
-
Transparency of Training Data: The complete recipe for 90M new multimodal data + 72M aesthetic data is fully disclosed.
-
Business-Friendly Agreement: MIT License + DeepSeek special authorization, zero threshold for enterprise-level applications.
This open stance is rapidly showing ecological effects: within 12 hours of the model’s release on Hugging Face, over 200 fine-tuning variants emerged, and GitHub stars surged at a rate of three per minute. While other vendors are still using API interfaces to “fence off” their territory, DeepSeek has already built its technological moat using an open-source ecosystem.
5. Future Outlook: The Fantastical Drift from Worker to Creator
Janus-Pro’s emergence marks a breakthrough for Chinese AI companies in achieving architectural definition rights in the core field of multimodality: from follower to leader. The GenEval score of 0.80 is not just a numerical surpass but a contest for the discourse power of the evaluation system. From application innovation to foundational innovation: 25 core technology patents such as the Rectified Flow generation framework have built an unfathomable technological barrier. From single-point breakthroughs to systemic capabilities: the smooth expansion from 1.5B to 7B verifies the foresight of the architecture, paving the way for the trillion-parameter era. What Janus-Pro showcases is merely the tip of DeepSeek’s ambition:
-
Multimodal Cognitive Flywheel: The closed-loop interaction of understanding and generation is giving rise to AI systems with self-iterative capabilities.
-
Industrial-Level Application Explosion: From medical image reconstruction to industrial design generation, tests show its error rate is 62% lower than SD3 in professional fields.
-
AGI Path Reconstruction: When a single model masters both language logic and visual creativity, we may be witnessing the embryonic form of general artificial intelligence.
On this magical night, we witnessed: the US stock market plummeting more thrilling than a roller coaster, AI competition more intense than the college entrance examination, and the Chinese team proving with strength: when the technology tree is fully developed, even Newton’s coffin board cannot hold back. On this turbulent night in the US stock market, DeepSeek announces to the world with Janus-Pro: the next chapter of the AI revolution is destined to be written by the Chinese team. This is not a coincidental explosion but the inevitable result of deep accumulation—while others chase after capital market fluctuations, true innovators are always silently rewriting the rules of the game.
Witness history? The portal is here:
This code repository is licensed under the MIT License.
https://github.com/deepseek-ai/Janus