Meta launches Muse Spark 1.2

Meta on August 5 released Muse Code, a terminal-based coding agent in beta, alongside Muse Spark 1.2, a new coding-focused model update. The company said the pair is designed for software engineering tasks across large repositories, including planning changes, writing code and validating results. Muse Code is available for macOS and Linux, according to Meta’s research blog. VentureBeat said the launch puts Meta into more direct competition with Anthropic’s Claude Code and OpenAI’s Codex in AI coding tools. ### What exactly did Meta ship? Meta said Muse Code is a terminal coding agent powered by Muse Spark 1.2, its newest model. In its product post, the company described Muse Spark 1.2 as a coding-focused update to Muse Spark 1.1, with improvements in code generation, debugging, codebase understanding and end-to-end developer workflows. (research.meta.ai) The August 5 post said Muse Code is aimed at “complex software engineering tasks across large repositories.” Meta said the tool can plan changes, write code and validate results, and that it was co-trained with Muse Spark 1.2 so the model and harness perform best together. ### Why are the “persistent background agents” the main feature? (research.meta.ai) Meta said Muse Code runs a main agent plus “async background agents” that stay active throughout a session instead of being created for one task at a time. The company said those agents help avoid repeated information gathering, can carry out next steps on their own and decide when to report back to the main agent. (research.meta.ai) VentureBeat reported that this is a different architecture from rival coding harnesses that often spawn helper agents fresh for each task. The publication said Meta’s pitch is lower latency and less need for developers to keep re-steering the system on multi-step work. (research.meta.ai) ### How does Meta say the system handles long jobs and failures? Meta said Muse Code uses a local event log that records every model call, tool run, approval and edit. The company said that log acts as a single source of truth and makes the runtime “replay-exact and restart-safe,” allowing the agent to resume where it stopped after a crash. (venturebeat.com) The same post said Muse Code can dispatch parallel sub-agents for larger jobs. VentureBeat reported that those sub-agents work in isolated git worktrees so the developer’s working copy is not touched directly. ### What did Meta say about how Muse Spark 1.2 was trained? Meta said Muse Spark 1.2 was trained with more compute on coding tasks and with a broader set of training environments than Muse Spark 1.1. (research.meta.ai) The company also said the model was extensively trained on long-horizon coding tasks, including whole-repository generation, large end-to-end projects and “auto-research.” Meta’s evaluation methodology page says the company compares Muse Spark models against Gemini 3.1 Pro, GPT 5.4, Claude Opus 4.6 and Grok 4.2 across reasoning, multimodal, coding and tool-use benchmarks. The page also says Meta sometimes uses self-reported third-party results for coding and agentic benchmarks when those are available. (research.meta.ai) ### How does this fit Meta’s broader AI strategy? Meta has been extending the Muse Spark line beyond consumer surfaces. Meta’s earlier materials said Muse Spark 1.1 was also made available to external developers through an API with tool and function calling, while the company’s June release notes for AI glasses said Muse Spark powers next-generation Meta AI features on those devices. (ai.meta.com) VentureBeat reported that Muse Code is also notable because it is proprietary, unlike the Llama family Meta had centered much of its developer story around. The publication said the launch gives Meta a fuller product in a market where coding agents have become a primary workflow for many developers. (ai.meta.com) ### Where can developers try it now? Meta said developers can install Muse Code on macOS or Linux from its developer site. VentureBeat reported that users need a Meta account and billing details to begin. Meta’s August 5 post also links to its evaluation materials for Muse Spark 1.2, which the company said provide more detail on performance. (research.meta.ai) (venturebeat.com)