⚙️Qwen2.5-1M: Deploy Your Own Qwen Version

🗞 News of the day

🚀 DeepSeek Cyberattack

Just last week, DeepSeek launched its free AI assistant, and it's already topping charts as the #1 free app on Apple's App Store in the U.S.! But with great popularity comes, well, a few hiccups. On Monday, DeepSeek faced a cyberattack that forced the company to temporarily limit new user registrations.

Despite the chaos, DeepSeek managed to resolve issues with its app and website, proving they're not just a flash in the pan. Their AI assistant, powered by the DeepSeek-V3 model, is turning heads with its performance, rivaling even the most advanced closed-source models out there. And get this—it's doing all this with less data and at a fraction of the cost of other AI giants.

DeepSeek's rapid rise is challenging the status quo, stirring up debates about U.S. tech export controls and the power dynamics in the AI world. Turns out, the company reportedly trained its AI on Nvidia's H800 chips, spending less than $6 million—a figure that's got tech execs scratching their heads and reevaluating their strategies.

DeepSeek is the first Chinese AI startup to earn accolades from the U.S. tech industry, proving that innovation knows no borders. Their story is a testament to how quickly the AI landscape can shift, and how the next big breakthrough might come from unexpected places.

While tech stocks took a hit on Monday, with Nvidia and Oracle shares plummeting, DeepSeek's success is a reminder that the AI race is far from over. The real question is: who's next to disrupt the status quo?

🔖 Today's Cool Story

🚨 DeepSeek Strikes Again

DeepSeek just dropped an update AI model called Janus Pro.

• Release of Janus Pro: DeepSeek has unveiled an updated AI model named Janus-Pro, available in two versions: Janus-Pro-1B and Janus-Pro-7B. The company shared a research paper detailing the model's performance, aiming to position Janus Pro as a competitor to OpenAI's DALL-E. This release follows the debut of DeepSeek's model R1, which showcased advanced reasoning capabilities.

• Performance and Benchmarking: Janus-Pro reportedly outperforms competitors like DALL-E 3 and Stable Diffusion 3 Medium across various benchmarks. DeepSeek's models are said to have been trained at a lower cost than those in the U.S., potentially disrupting the current AI model landscape. The industry is closely watching these developments, as they could shift existing AI dynamics.

• Market Impact: DeepSeek's AI app has reached the top ranking in Apple's App Store for free apps, highlighting its market influence. The company's competitive pricing strategy could challenge traditional AI business models and affect profitability expectations. As a key player in the AI industry, DeepSeek's advancements may shape the future of AI research and applications, offering cost-effective solutions and increasing accessibility.

💡 AI Insight

🎯 Qwen2.5-1M: Deploy Your Own Qwen with Context Length up to 1M Tokens

The Qwen2.5-1M models are setting new standards in AI with their capacity to handle extremely long contexts, boasting the ability to process up to 1 million tokens.

• Innovative Model Features: The release includes the Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M models, both of which have been enhanced to manage substantial context lengths. The models are complemented by a new open-sourced inference framework based on vLLM, which guarantees efficient deployment and operation.

• Superior Performance: In tasks involving long contexts, the Qwen2.5-1M models outperform their predecessors, particularly when handling sequences longer than 64K tokens. The Qwen2.5-14B-Instruct-1M model significantly surpasses competitive models such as the Qwen2.5-Turbo and GPT-4o-mini across multiple datasets.

• Advanced Techniques: The models utilize several sophisticated techniques, including a progressive long-context training approach, dual chunk attention for managing large positional distances, and a sparse attention mechanism to enhance prefill phase speed. These innovations not only improve the models' efficiency and accuracy but also guarantee they are ready for real-world applications that demand handling of extensive data sequences.

🤖 Today on AI Robotics

🚨 Humanoid Robots Might Be Coming to Your Next Karaoke Night

Humanoid robots are stepping into the world of music, jamming alongside human musicians in an innovative study.

• Robots Join the Band: The study, published in PeerJ Computer Science, features a unique human-robot musical band. Humanoid robots Polaris, a drummer, and Oscar, a keyboardist, are making waves by collaborating with human musicians.

• Innovations in Performance: The study introduces advanced synchronization techniques, allowing robots to adapt to human musical cues naturally. Utilizing the Robot Operating System (ROS), these robots adjust and interpret musical notes in real-time, showcasing their potential for creative collaboration in live performances.

• Future of Humanoid Musicians: Researchers aim to refine synchronization algorithms to improve robots' timing and beat recognition. Future directions include developing robots capable of engaging audiences, improvising music dynamically, and exploring broader applications in education, therapy, and public engagement.

  The significance of this research lies in its potential to bridge the gap between technology and creativity, demonstrating the adaptability and artistry robots can bring to various sectors.