Meta Releases Muse Code Terminal Agent and Muse Spark 1.2 Model
Meta released Muse Code, a CLI terminal coding agent with persistent async background subagents and restart-safe local logging. Powered by the co-trained Muse Spark 1.2 model, it automates long-horizon tasks and Triton GPU kernel optimization.
Impact: High
Why it matters
Engineers can install a restart-safe terminal agent with specialized slash commands for repository-level tasks and custom GPU kernel tuning.
TL;DR
- 01Muse Code uses an append-only event log to ensure replay-exact session recovery after crashes.
- 02Async background subagents remain active across tasks, avoiding redundant context collection.
- 03Default slash commands include /plan, /grill for stress-testing, and /goal for long-running targets.
Key facts
- Optimization Horizon
- 1,000+ tool calls (up to 24 hours)
- Target GPU Architecture
- NVIDIA Hopper (Triton KDA / MLA kernels)
- Installation Method
- curl -fsSL https://dev.meta.ai/install.sh | bash
Terminal-First Agentic Architecture
Muse Code operates as a CLI application powered by Meta's Muse Spark 1.2 model. The agent uses a local event log that records every model call, tool execution, edit, and user approval in an append-only format, ensuring replay-exact session restoration after interruptions.
Slash Commands and Workflow Orchestration
Built-in workflow triggers include:
/plan: Generates approval-gated task plans before code execution./grill: Stress-tests generated plans against edge cases and failure modes./goal: Directs persistent subagents toward targeted objective completion.
Autonomous Triton Kernel Optimization
During evaluations on NVIDIA Hopper GPUs, Muse Spark 1.2 continuously generated, compiled, and benchmarked Triton GPU kernels over 1,000+ tool iterations spanning up to 24 hours. The model implemented chunk-parallel preparation kernels and sequential inter-chunk scans for KDA and MLA algorithms without external library imports.
Try it in 2 minutes
curl -fsSL https://dev.meta.ai/install.sh | bashbash
✓ When to use
- Running multi-step coding workflows across large repositories in the terminal.
- Iteratively writing and benchmarking low-level GPU kernels in Triton.
- Long-running tasks that require persistent subagent coordination and crash recovery.
✕ When NOT to use
- Quick single-prompt edits that do not justify background agent execution.
- Environments requiring verification without Meta account setup.
What to do today
- Install Muse Code via dev.meta.ai install script.
- Test /plan and /grill commands on a sub-module before committing large context runs.
What the community says
“And they are all TUI's installed via curl | bash.”
Sources