Meta launched Muse Code, a terminal-based agentic coding tool now in beta, alongside Muse Spark 1.2, a coding-specialized model update. Install it on macOS or Linux with one command: curl -fsSL https://dev.meta.ai/install.sh | bash. A Meta account and billing details are required before you can start. This puts Meta in direct competition with Anthropic's Claude Code and OpenAI's Codex, markets it largely ignored while distributing Llama as open-weight software. Muse Spark 1.2 is proprietary, though Zuckerberg hinted on X that open source may follow, writing 'I'll have more to share on that soon.'
The architecture has two bets worth understanding. First, Muse Code runs persistent async background agents that stay alive for the entire session rather than spinning up fresh per task, reducing redundant repository exploration and latency. For large jobs, it fans out to parallel sub-agents each working in an isolated git worktree, leaving the developer's working copy untouched. Zuckerberg said in testing it built six game features simultaneously with zero collisions. Second, every model call, tool run, and edit is written to a local event log before execution, making the runtime replay-exact and restart-safe. A 20-hour job interrupted mid-run resumes precisely where it stopped.
The benchmark numbers tell the real story and are worth reading in full. On Terminal-Bench 2.1, Muse Spark 1.2 inside Muse Code scored 82.9%, beating OpenAI's GPT-5.6 Terra in Codex at 81.8% and xAI's Grok 4.5 in Grok Build at 81.6%, but trailing Anthropic's Opus 5 in Claude Code at 86.7%. On DeepSWE 1.1 it posted 59.3%, third behind Opus 5 at 65.0% and GPT-5.6 Terra at 64.8%. On Meta's own internal coding benchmark, Muse Spark 1.2 scored 70.6% while Opus 5 scored 79.4%, an unusually candid admission that Anthropic's model wins even the test Meta designed itself. The generational gains over Muse Spark 1.1 are real, 6.7 points on Terminal-Bench and 6.3 on DeepSWE, but the gap to close is documented and specific.
[READ ORIGINAL →]