[“Manus 2.0 is released: the Cascade framework reduces token usage by 23.2% and launches a desktop app”]
Manus 2.0 has been officially released, with the Cue personal Agent platform also going live at the same time. This update introduces a proprietary Cascade framework, using a lightweight distributed architecture that only calls on specialized capabilities when necessary. According to official data, token consumption has been reduced by 23.2%, task execution speed has increased by 28.2%, and overall costs have been lowered by 32%. The new architecture supports cloud deployment and features capabilities such as “continue executing online after closing the cover,” task self-healing, and long-horizon planning. In addition, the Manus Studio desktop app (Mac/Windows) introduces the “Give Manus My Computer” feature, allowing AI to access the local system, including permissioned terminal commands and local folders, directly dispatching local GPU computing power to enable codebase refactoring and local data cleanup. The Cue platform also supports users freely assembling their own dedicated personal Agent teams to handle everyday tasks such as itinerary planning and account management. This upgrade marks the expansion of AI Agents from a web sandbox to local systems and a distributed ecosystem, which may affect the cost structure of AI toolchains and application scenarios. Key areas to watch next: the stability of the Cascade framework in real multi-task scenarios, and the impact of the local GPU scheduling feature on existing AI workflows.
Manus 2.0 has been officially released, with the Cue personal Agent platform also going live at the same time. This update introduces a proprietary Cascade framework, using a lightweight distributed architecture that only calls on specialized capabilities when necessary. According to official data, token consumption has been reduced by 23.2%, task execution speed has increased by 28.2%, and overall costs have been lowered by 32%. The new architecture supports cloud deployment and features capabilities such as “continue executing online after closing the cover,” task self-healing, and long-horizon planning. In addition, the Manus Studio desktop app (Mac/Windows) introduces the “Give Manus My Computer” feature, allowing AI to access the local system, including permissioned terminal commands and local folders, directly dispatching local GPU computing power to enable codebase refactoring and local data cleanup. The Cue platform also supports users freely assembling their own dedicated personal Agent teams to handle everyday tasks such as itinerary planning and account management. This upgrade marks the expansion of AI Agents from a web sandbox to local systems and a distributed ecosystem, which may affect the cost structure of AI toolchains and application scenarios. Key areas to watch next: the stability of the Cascade framework in real multi-task scenarios, and the impact of the local GPU scheduling feature on existing AI workflows.