Meta ships Muse Spark 1.2

Meta released Muse Code in beta and Muse Spark 1.2 on August 5, describing the pair as its latest push into AI-assisted software development. Muse Code is a terminal-based coding agent for macOS and Linux, while Muse Spark 1.2 is the model Meta says was tuned for code generation, debugging, codebase understanding and end-to-end developer workflows. Meta said the products were built to handle longer-running engineering tasks across large repositories rather than one-off code completions. VentureBeat described the launch as Meta entering the “AI coding wars,” placing the company more directly against coding tools from rivals already competing for developers and enterprise budgets. Meta framed the release more narrowly, saying Muse Code can plan changes, write code and validate results, and that larger models are still in development. (research.meta.ai) ### What exactly did Meta ship? Meta said Muse Code is a terminal coding agent powered by Muse Spark 1.2, its newest model. The company said the agent is designed for “complex software engineering tasks across large repositories,” including planning, implementation and validation. August 5 marked the first public beta for Muse Code, and Meta’s install instructions specify macOS and Linux. (venturebeat.com) The same announcement said Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with improvements in code generation, complex debugging, codebase understanding and broader developer workflows. ### Why are the background agents the key detail? (research.meta.ai) Meta said Muse Code uses a main agent plus “async background agents” that remain active throughout a session instead of being spun up for each task. According to the company, those subagents gather information, carry out next steps and decide when to report back to the main agent, reducing redundant work and cutting the amount of steering needed on multi-step jobs. (research.meta.ai) The company said that persistence matters because coding jobs often stretch across long sessions, multiple files and repeated checks. Meta also said Muse Code keeps a local event log in which every model call, tool run, approval and edit is appended, allowing the runtime to resume after a crash at the exact point where it stopped. ### How is Muse Spark 1.2 different from earlier Muse models? (research.meta.ai) Meta introduced the original Muse Spark in April as the model family behind Meta AI across its app, website and, over time, other products including WhatsApp, Instagram, Facebook, Messenger and AI glasses. In July, Meta said Muse Spark 1.1 also powered new planning and task-follow-through features in Meta AI, including recurring briefings, research tasks and app-connected actions. (research.meta.ai) Muse Spark 1.2 extends that line with a stronger coding focus. Meta said it increased training compute on coding tasks, expanded the diversity of training environments and trained the model extensively on long-horizon coding work such as whole-repository generation, end-to-end projects and “auto-research.” ### Did Meta build the model and the agent together? (about.fb.com) Meta said it “co-trained” Muse Spark 1.2 with Muse Code so the model would perform best when paired with the agent. The company said that work included harness trajectories, optimizations for goals, compaction and subagents, and integration of the Muse Code toolset. That approach suggests Meta is not treating the model and the coding interface as separate products. (research.meta.ai) VentureBeat said the release underscores a broader contest around integrated coding assistants, workflow orchestration and developer experience, rather than simple autocomplete. ### What does Meta say developers can actually do with it? Meta said Muse Code ships with built-in commands including `/plan`, which turns a task into an approval-gated plan, `/grill`, which stress-tests that plan, and `/goal`, which works toward completing an objective. The company’s example described a user dropping a fly-through home video into the terminal and having Muse Code produce a marketing and booking page. (venturebeat.com) Meta’s August 5 post also directs users to a separate evaluation report for Muse Spark 1.2. For now, the immediate next step is distribution: Muse Code is available in beta for macOS and Linux, and Meta said larger and “much more capable” models are on the way. (research.meta.ai)