According to CoinMeta, REDnote's dots laboratory has open-sourced dots3-note, which is the first open-weight model in the dots3 series. This model has a total of 280 billion parameters, with 16 billion parameters activated at a time, and it supports 512k contexts. It is capable of processing text, images, videos, and audio, with a focus on long-term agent. The model can explore rules, call tools, write code in unfamiliar environments, and store important information in its memory, adjusting its actions based on new circumstances. Its training method tempo is designed to address the issue of delayed feedback for long tasks by breaking down tasks that take over a dozen hours into multiple stages, allowing the model to self-evaluate midway and then continue with reinforcement learning. Official experiments show that ordinary grpo tends to stagnate in the later stages, while tempo can still continue to improve. The team has also open-sourced vibesearchbench and vibelifebench; the latter includes 200 tasks, 22 simulation services, and 288 tool interfaces, but the current average performance of the strongest model is only 32.5. There is still much work to be done on long-term agent; there is still a significant distance to go before truly completing long-term tasks.