Meta has released Muse Code in beta, a terminal coding agent for complex software engineering work across large repositories. Available on macOS and Linux, it can plan changes, write code, and validate results. The agent runs on Muse Spark 1.2, a coding-focused model now offered through Meta Model API with expanded global access.
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 runtime pairs a main agent loop with persistent asynchronous background agents that stay active throughout a session. They gather context, take next steps, and decide when to report back, avoiding repeated information collection during multi-stage tasks. Muse Code can coordinate multiple agents on a single job to reduce latency and manual steering.
Every model call, tool run, approval and edit is appended to a local event log. Meta says this makes work replay-exact and restart-safe, so an interrupted task can resume from the same point after a crash. Built-in skills can create an approval-gated plan, stress-test it and continue toward a defined goal.
To install Muse Code, run
Muse Spark 1.2 builds on version 1.1 with stronger code generation, complex debugging, codebase understanding and end-to-end development workflows. Meta increased training compute for coding and expanded the range of training environments. It co-trained the model with Muse Code using rejection-sampled harness trajectories and recipe optimizations for goals, context compaction and subagent work.
Training covered long-horizon jobs including whole-repository generation, large projects and automated research. Meta also used Muse Spark 1.1 to create demanding coding environments and instruction templates, then grade candidate solutions to build training data for its successor. The company says this helped the new model follow complex requirements more precisely.
In one case study, the system made more than 1,000 tool calls over runs lasting up to 24 hours while optimizing GPU kernels for NVIDIA Hopper hardware. It wrote, compiled and profiled Triton implementations, revising them against a baseline without importing third-party kernel libraries.
The release shows Meta coupling a coding model with the runtime used to train and operate it, framing the pair as one developer system. Meta says more harness features and larger, more capable models are still ahead.
