China's DeepSeek Quietly Drops New AI Model

The AI race is heating up with groundbreaking developments reshaping the industry. Chinese startup DeepSeek is shaking things up with its open-source DeepSeek-V3-0324, a 685-billion-parameter model that runs on consumer hardware. Meanwhile, Tencent’s Hunyuan-T1 is pushing speed and reasoning limits, outperforming top-tier models like GPT-4.5 with its Hybrid-Mamba-Transformer MoE architecture. On the platform front, Perplexity envisions an AI-driven TikTok, prioritizing transparency, citations, and enhanced search functionality. From open-source breakthroughs to ultra-fast AI models and a potential social media overhaul, these updates highlight the next wave of AI evolution.

Let’s explore!

DeepSeek Launch V3- 0324 Open-Source AI Model

Chinese AI startup DeepSeek has launched its latest large language model, DeepSeek-V3-0324, which is making waves in the AI industry for its open-source MIT license and ability to run on consumer-grade hardware like Apple’s Mac Studio with an M3 Ultra chip. This 685-billion-parameter model employs a mixture-of-experts (MoE) architecture, activating only relevant parameters for tasks, drastically improving efficiency. It also incorporates innovations like Multi-Head Latent Attention (MLA) and Multi-Token Prediction (MTP), boosting speed by 80%. Unlike Western companies like OpenAI, which keep models proprietary, DeepSeek’s open-source approach democratizes AI access, enabling developers and researchers to use cutting-edge technology without high costs. The model’s release strategy—quietly uploading it to Hugging Face without fanfare—contrasts with the marketing-heavy launches of Western counterparts. Early testers report significant performance improvements over competitors like Anthropic’s Claude Sonnet 3.5, positioning DeepSeek-V3-0324 as a leader in non-reasoning AI models. This open-source strategy reflects China’s emphasis on efficiency due to limited access to advanced hardware, challenging Silicon Valley’s closed ecosystem and accelerating innovation globally.

Chinese Startup Tencent Launch Hunyuan-T1 AI Model

Tencent's new AI model, Hunyuan-T1, is making waves in the AI community with benchmarks suggesting it outperforms DeepSeek R1 and even OpenAI's GPT-4.5. Boasting a generation speed of 60-80 tokens per second, this model employs a Hybrid-Mamba-Transformer MoE architecture for superior reasoning and high-quality output. Hunyuan-T1 activates 52 billion parameters through dynamic expert routing, utilizes Cross-Layer Attention (CLA) to reduce GPU memory consumption, and maintains FP16 precision while doubling inference speed through FP8 Quantization. Trained on 4.8 trillion tokens of multilingual data, with 65% being Chinese, it features a 256K context window and synthetic data augmentation. Access is currently primarily through the Tencent Yuanbao platform, requiring a Chinese phone number, which limits global accessibility despite its impressive capabilities.

Perplexity's Vision: Rebuilding TikTok as a Transparent, AI-Enhanced Platform

Perplexity, aiming to leverage its AI expertise, proposes rebuilding TikTok into a more transparent and useful platform. The vision includes developing a transparent, open-source algorithm to replace TikTok's current "black box" system, ensuring neutrality and trust. By integrating Perplexity's AI infrastructure, powered by Nvidia Dynamo, TikTok's recommendation models could scale significantly with faster inference speeds. Furthermore, Perplexity plans to enhance trust by adding citations to TikTok videos and developing community note features. The integration of Perplexity's answer engine would improve TikTok's search functionality, while AI-driven content enrichment, including translation and annotation, would expand creators' reach. This rebuilt TikTok would prioritize user satisfaction and informational value, creating a platform for both creativity and knowledge discovery.

Hand Picked Video

In this video, we'll look at China just dropped QwQ-32B, and it’s making waves in the AI world! This model is being called the smallest yet most powerful reasoning model out there.

Top AI Products from this week

• BASE44 2.0 - Base44 is the only all-in-one platform that empowers anyone to create software. From tools and personal apps - to complete online businesses—without writing a single line of code (unless you want to).

• LiftmyCV - Tired of filling out the same job application forms over and over? Exhausted from browsing dozens of job search websites? I've got you covered! Meet LiftmyCV – an AI agent that finds job openings and auto-applies on your behalf. No subscription required.

• Lamatic 2.0 - A fully managed PaaS with Agent IDE, integrated vector stores, and seamless connections to apps, data sources, and leading AI models — empowering you to rapidly build, test, and deploy high-performance AI Agents at the Serverless Edge.

• Krea Video Training - Krea Video Training is the fine-tune Krea's AI video model with your own images/videos to create custom styles, motions, and objects. Control the AI, don't just prompt it!

• Interview Code Ace - The ultimate, invisible AI powered tool to help you succeed in technical interviews. Practice, prepare, and perform with confidence.

• The "think" tool from Claude - With the "think" tool, we're giving Claude the ability to include an additional thinking step—complete with its own designated space—as part of getting to its final answer.

This week in AI

• ChatGPT Emotional Impact Study - OpenAI's study with MIT found ChatGPT users emotionally engage less; women socialized less, and cross-gender voice use increased loneliness. More research is needed.

• Grok 3 Adds New Features - xAI's Grok 3 now offers "DeeperSearch" for detailed queries and text-based image editing. It enhances research capabilities and supports photorealistic image modifications.

• Neo-1 - VantAI's Neo-1 is a foundation model unifying structure prediction & molecular design. It enables programmable design of biomolecules, accelerating drug discovery by integrating experimental data.

• AI Detects Cancer Accurately - A new AI model, ECgMLP, detects endometrial cancer with 99.26% accuracy, surpassing existing methods. It shows promise for diagnosing other cancers as well.