Meta’s Muse Code uses Muse Spark 1.2 in a 1,000+ tool-call GPU optimization test

Meta has launched Muse Code, a terminal-based coding agent that completed more than 1,000 tool calls during a test lasting up to 24 hours, as the company pushes its latest coding model toward longer and more complex software tasks. Muse Code is powered by Muse Spark 1.2, Meta’s latest coding-focused model, and is designed to work across large software repositories. It can plan changes, write code, and validate the results rather than stopping after generating a code snippet. The system can also coordinate multiple persistent background agents during a session. Instead of creating a new helper for every individual task, these agents remain active, carry out follow-up work and decide when to send information back to the main agent. Meta says this approach is intended to reduce repeated information gathering, latency and the amount of user intervention required for multi-step engineering work. Coding tasks that run longer One of the more demanding tests involved GPU kernel optimization. Muse Spark 1.2 worked inside Muse Code to write, compile, profile and repeatedly improve GPU kernels over more than 1,000 tool calls, with individual runs lasting up to 24 hours. Releasing Muse Code in beta today. It's a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results. Powered by Muse Spark 1.2, a coding-focused model update. pic.twitter.com/xqavk41w6v— Mark Zuckerberg (@finkd) August 5, 2026 The tests used KDA and MLA kernels running on Nvidia Hopper GPUs. The system had to improve performance against a supplied baseline implementation rather than simply import an existing third-party kernel library. For KDA, the model combined a chunk-parallel preparation kernel with a sequential inter-chunk scan. It also applied techniques such as fusion and tiling, along with optimizations specific to the algorithm. The test is meant to show whether a coding system can maintain direction across a long sequence of engineering decisions instead of producing an answer in a single pass. Muse Code also uses a local event log that records model calls, tool runs, approvals, and edits. This creates what Meta describes as a replay-exact and restart-safe runtime. If the system crashes during a lengthy task, it can resume from where it stopped rather than starting again. That capability could be particularly useful for coding jobs that involve repeated compilation, testing, and optimization, where a failure late in the process could otherwise erase hours of work. Persistent agents expand coding Muse Code includes built-in tools for managing these longer workflows. Its /plan function creates an approval-gated plan, /grill can stress-test that plan and /goal works toward completing a specified objective. Meta is also using Muse Spark 1.2 for broader software engineering tasks, including whole-repository generation, complex debugging, and long-running projects. The model was co-trained with Muse Code, including training on agent trajectories and the tool environment itself. Another demonstration shows the system taking an MP4 fly-through video of a home and producing a vacation-home marketing and booking page through the terminal. The example illustrates how the coding agent can work from inputs beyond conventional source code.Muse Spark 1.2 is available through Muse Code and Meta’s Model API, with expanded global access. Meta said it plans to add more harness features and larger models.Recommended ArticlesGet the latest in engineering, tech, space & science - delivered daily to your inbox.With over a decade-long career in journalism, Neetika Walter has worked with The Economic Times, ANI, and Hindustan Times, covering politics, business, technology, and the clean energy sector. Passionate about contemporary culture, books, poetry, and storytelling, she brings depth and insight to her writing. When she isn’t chasing stories, she’s likely lost in a book or enjoying the company of her dogs.