Descriptions:
Fahd Mirza reviews Dot3 Note Preview, the debut model in a new family from Dot Studio — the AI research arm of Xiaohongshu (Red Note), the Chinese social platform often described as Instagram meets Pinterest. The model is a 280-billion-parameter mixture-of-experts with only 16 billion active parameters, accepting multimodal input across text, image, video, and audio, with up to 512k token context. Benchmark charts show it leading its weight class on ARC-AGI-3 and outperforming significantly larger dense models and MoE competitors like Gemini K3 and GPT-3.5 Omni Plus across reasoning, coding, and long-horizon agent tasks.
The video is structured around live agentic tests. In the first, the model is given real AWS credentials and a single goal: provision an EC2 instance without any hand-holding. It identifies the AWS CLI, executes commands, configures a security group with SSH port 22, and launches a named instance (“dot-three-note-demo”) — completing the task in one pass before the host terminates the instance. The second test asks the model to analyze a UI screenshot of a restaurant interface and rebuild it as a fully themed animated web app. The output correctly replicates color theming per food category, adds animated flame effects, and switches simulation titles by tab — details the host highlights as evidence of genuine visual grounding.
A third test probes complex role-based decision-making under emotional constraints. Dot3 Note is a notable entry from a Chinese lab that has received less Western coverage than DeepSeek or Qwen, and this video offers a practical first look at its agentic capabilities.
📺 Source: Fahd Mirza · Published August 20, 2026
🏷️ Format: Hands On Build
