According to CoinMeta, Meta has released the programming model Muse Spark 1.2 and simultaneously opened the beta version of terminal programming Agent Muse Code. This model is trained together with Agent to better adapt to each other's work processes. Muse Code is capable of planning modifications, writing code, and verifying results within large codebases. When faced with larger tasks, it breaks them down into multiple sub-tasks to be processed in parallel by different Agent in isolated workspaces, without directly altering the user's current workspace. Every time the model is called, a tool is executed, or a file is modified, it is first recorded in a local event log. Even if it crashes midway, Muse Code can continue from where it left off, eliminating the need to re-enter all the information prompt. In addition, Meta also enabled Muse Spark 1.2 to optimize NVIDIA Hopper's GPU kernel for 24 consecutive hours, during which more than 1000 tool calls were made, with further performance improvements still being sought in the latter half of this process.