Meta launches Muse Code and Spark

Meta launched Muse Code in beta on August 5 as its first coding agent, pairing the terminal-based tool with a new model, Muse Spark 1.2, as it pushes deeper into software-development tools. Meta said Muse Code is built for large repositories and can plan changes, write code and validate results, putting it into direct competition with coding products from Anthropic and OpenAI. (research.meta.ai) The new release stands out less for basic code completion than for how it runs. (cnbc.com) Meta said Muse Code uses a main agent plus persistent asynchronous background agents that stay active through a session, rather than spinning up for each task. (research.meta.ai) The company said that design is meant to reduce repeated information gathering and help with long, multi-step work. ### What exactly did Meta ship? Meta AI Research said Muse Code is a terminal coding agent for macOS and Linux, and described Muse Spark 1.2 as a coding-focused update to Muse Spark 1.1. The company said the model improves code generation, debugging, codebase understanding and end-to-end developer workflows. (research.meta.ai) Meta also said Muse Code keeps a local event log that records every model call, tool run, approval and edit. Meta said that log makes the runtime “replay-exact and restart-safe,” allowing the agent to resume after a crash. (research.meta.ai) ### Why are the background agents getting attention? Meta said the background agents remain active for the whole session and can decide when to report back to the main agent. The company said that persistence reduces latency and the amount of steering needed on harder tasks. VentureBeat described that architecture as a move toward “persistent async background agents” for repository work, rather than one-shot assistant responses. (research.meta.ai) That framing matters because the new tools are aimed at longer-running engineering jobs, not just inline suggestions. ### Where does pricing enter the story? Alexandr Wang, Meta’s AI chief, told CNBC that developers can access Muse Code through a pay-as-you-go option with pricing similar to Muse Spark 1.1: $1.25 per million input tokens and $4.25 per million output tokens. Wang also said Muse Code has a contributor tier that is “more than 10 times cheaper” than the pay-as-you-go tier. (research.meta.ai) CNBC reported that the lower-cost tier requires developers to “opt-in to help improve the model,” in Wang’s words. VentureBeat reported the same trade-off more directly, saying the default on-ramp sends developers’ code and prompts into Meta’s training pipeline unless customers move to standard pricing. (venturebeat.com) ### Why does that create a governance problem for some users? VentureBeat reported that the default data-sharing setup could be a problem for enterprises with proprietary codebases. That is a product-design fact with legal and policy implications that will depend on each customer’s rules, but the immediate issue is straightforward: teams have to make an active pricing and data-use choice before using the tool on sensitive repositories. (cnbc.com) Meta has already faced scrutiny over AI training data in Europe. Meta said in a 2023 privacy post that its generative AI models are trained on public online information, licensed data and information from Meta’s products and services, while outside reporting has documented opt-out disputes around European user data. (cnbc.com) That history is why data-handling terms around a coding agent are likely to draw attention from compliance, privacy and security teams. ### How is Meta positioning this against rivals? Meta said Muse Spark 1.2 was co-trained with Muse Code to improve coding performance and usability when the two are used together. (venturebeat.com) CNBC reported that Wang is leading the release through Meta Superintelligence Labs as Meta tries to challenge Anthropic and OpenAI in AI developer tools. Meta’s August 5 launch post says Muse Code is available now in beta, with installation through a one-command setup on macOS or Linux. The company also linked an evaluation report and said larger, more capable models are still to come. (research.meta.ai) (about.fb.com)