A pinnacle showdown in the AI realm has just concluded, with the domestic AI model "Flash" ascending to the top of the global AI model rankings, surpassing the previously dominant GPT-4o!
From one of the "Big Six," Zero One Everything, this new "Flash" model has made a remarkable debut in the rankings, breaking into the top six of the overall list, and tying for third place in the mathematics sub-ranking alongside international giants, with top rankings in other sub-rankings as well. Notably, Flash's performance is nearly on par with Elon Musk's latest xAI model, Grok-2-08-13, significantly outperforming other competitors such as GPT-4o-2024-05-13, GPT-4o-mini-2024-07-18, and Claude3.5Sonnet.
This ranking was determined through blind tests and votes from over ten million global users, ensuring unquestionable fairness. The rise of Flash also signifies that domestic AI models now have a stronger voice on the global stage.
More importantly, the selection criteria for this ranking underwent significant adjustments, placing greater emphasis on the real-world problem-solving capabilities of AI models, rather than overemphasizing response length and style. Under these new rules, Flash maintained its leading position, competing in the same tier as GPT-4o and Grok-2, fully demonstrating its formidable strength.
The technological innovations behind the Flash model are also impressive. It employs a mixed-expert model architecture, optimizing the mixed attention mechanism, which significantly reduces inference costs when processing long sequence data. Additionally, Flash introduces a cross-layer attention mechanism, making information sharing between different layers more efficient.
Interestingly, Flash can dynamically select the most appropriate combination of expert networks based on task difficulty. During the training phase, all expert networks are activated, but during the inference phase, Flash selectively activates relevant experts based on specific tasks, ensuring it can handle various challenges with ease.
Dr. Kai-Fu Lee, founder and CEO of Zero One Everything, emphasized at the launch event that the Flash model employs various strategies during the pre-training phase to ensure not only speed but also the acquisition of richer knowledge. Notably, Flash's pricing strategy is also highly competitive, costing only 0.99 yuan per million tokens, offering excellent value for money.
Regarding the gap with top international models, Dr. Lee stated that Flash's outstanding performance has narrowed this gap to just five months. This is not only a technological advancement but also a reflection of confidence in the domestic AI industry.
With the emergence of the Flash model, domestic AI is showcasing its exceptional strength and potential to the world. We look forward to seeing how it will continue to break new ground and lead the trend in the future AI field.