Meta has entered the competitive field of AI coding tools with the beta release of Muse Code, a terminal-based agent, and Muse Spark 1.2, a specialized update to its Muse model family. The tools are designed to handle full software engineering tasks across large repositories, including planning changes, writing code, and validating results. Developers can install Muse Code on macOS or Linux with a single command, though access requires a Meta account and billing details before any work begins.
The agent’s core design relies on persistent background processes that stay active for an entire session rather than launching fresh helpers for each request. This approach aims to reduce repeated exploration of the same codebase and cut down on latency. For larger jobs the system can spin up parallel sub-agents inside isolated git worktrees so the main working copy remains untouched. Every action is recorded in a local event log, allowing the process to resume exactly where it left off after a crash. These choices address common frustrations with earlier coding agents that often lost context or required constant supervision.
Muse Spark 1.2 was trained specifically alongside the new harness, using trajectories generated by the previous version and a self-improvement loop that created harder coding environments. Benchmarks place the combination in a strong but clear second tier. On one terminal-focused test it scored just behind the leading Anthropic system while edging out several other models; on others it trailed more noticeably. A long-running demonstration showed the agent optimizing GPU kernels over 24 hours and more than a thousand tool calls, producing non-obvious improvements without relying on existing libraries. Gains over the prior Muse version are measurable, though some of the reported progress may stem from the new harness rather than the model alone.
Pricing follows a two-tier structure that reflects Meta’s current strategy. The standard rate sits in the middle of the market and comes with a commitment not to train on user data. A far cheaper contributor tier undercuts nearly every rival, but only in exchange for permission to use prompts and completions for future training. Rate limits on the low-cost option are tight, signaling it is intended for individuals and experiments rather than production. Enterprises with proprietary code will need to weigh the savings against the data tradeoff.
The release marks a clear departure from Meta’s earlier emphasis on open-weight models. After years of positioning free, downloadable weights as a path to widespread adoption, the company now ships a proprietary agent and model with no open weights or self-hosting option. Rivals have moved in the opposite direction by open-sourcing their own coding tools. Meta’s pivot places it closer to the closed approach of other major labs while using discounted access to gather training data at scale. Whether the combination of persistent agents, auditability, and aggressive pricing can close the performance gap on real-world repositories will determine how seriously developers treat the new entrant.
