Red Cannon Open Sources Dots3-Note: Enabling Agents to Take Notes on the Fly, Specializing in Ultra-Long Tasks

According to Vision One monitoring, RedBook's dots lab has open-sourced the dots3-note preview, which is the first open-weight model in the dots3 series. With a total of 280B parameters, 16B per activation, supporting 512K contexts, it can process text, images, videos, and audio.

Its main focus is on Long-range Agents. The model can explore rules, invoke tools, write code in unfamiliar environments, store important information in memory, and adjust actions based on new circumstances.

The training method TEMPO specifically addresses the issue of late feedback in long tasks. It breaks down tasks that last for hours into multiple stages, allowing the model to self-assess midway through and then continue reinforcement learning. In official experiments, while regular GRPO tends to plateau later on, TEMPO can still show improvement.

The team has also open-sourced VibeSearchBench and VibeLifeBench. The latter includes 200 tasks, 22 simulated services, and 288 tool interfaces, with the current strongest model avg@3 only reaching 32.5. Long-range Agents still have a long way to go to truly accomplish tasks in the long term.