Xiaohongshu — the Chinese social platform also known as Little Red Book — has open-sourced dots3-note preview, the first open-weights model in its dots3 series. It's a hefty one: 280 billion total parameters, 16 billion active per forward pass, a 512K-token context window, and support for text, images, video, and audio.
The model is aimed at long-horizon agents. In unfamiliar environments, it can explore to learn the rules, call tools, write code, save important information to memory, and adjust its actions as new situations arise.
To train it, the team came up with TEMPO, a method that addresses the delayed-feedback problem in long tasks. Instead of waiting for a reward signal after hours of work, TEMPO splits the job into stages, lets the model assess itself midway, and keeps the reinforcement-learning loop going. In official experiments, standard GRPO tends to stall in the later stages — TEMPO keeps improving.
The lab also released VibeSearchBench and VibeLifeBench. VibeLifeBench includes 200 tasks, 22 simulated services, and 288 tool interfaces. The current strongest model only reaches an avg@3 of 32.5. Long-horizon agents are still far from actually getting long-term jobs done.
The team shared more in an announcement from Xiaohongshu's dots lab.