Muse Code and Muse Spark 1.2
Announcing Muse Code and Muse Spark 1.2
Published August 5, 2026 | Reading Time: ~4 minutes
Meta AI Research is proud to unveil Muse Code (beta), a sophisticated terminal-based coding agent driven by our latest model, Muse Spark 1.2.
"This release represents our next leap toward the frontier of AI, paving the way for even larger and more capable models in the near future."
🚀 Getting Started
To install Muse Code on your Linux or macOS system, execute the following command in your terminal:
curl -fsSL https://dev.meta.ai/install.sh | bash
🛠️ Muse Code: The Agentic Framework
Muse Code is designed to handle intricate software engineering challenges across expansive repositories. It doesn't just write snippets; it manages the entire lifecycle of a change:
- Strategic Planning: Mapping out necessary modifications.
- Implementation: Writing the actual source code.
- Verification: Validating that the results meet the requirements.
The Agent Architecture
Unlike traditional agents that spawn temporary workers, Muse Code utilizes a primary agent loop supported by persistent async background agents.
These specialized agents remain active throughout the session, which minimizes redundant data collection and reduces the need for constant user steering on complex, multi-step workflows.
Runtime Reliability & Control
The system relies on a local event log—a comprehensive record of every tool execution, model call, and edit. This creates a "single source of truth," ensuring the runtime is:
- Replay-exact: Every step can be mirrored.
- Restart-safe: If a crash occurs, the agent resumes exactly where it left off.
Essential Commands
| Command | Function |
|---|---|
/plan | Converts a request into a plan requiring user approval. |
/grill | Rigorously stress-tests a plan to find weaknesses. |
/goal | Directs the agent to work until a specific objective is achieved. |
Example Use Case: A user provides an .mp4 fly-through video of a residence. Muse Code analyzes the visual data and automatically generates a high-end, visually appealing marketing and booking website for that vacation home.
🧠 Muse Spark 1.2: The Engine
Muse Spark 1.1 Muse Spark 1.2
This update focuses heavily on coding proficiency, enhancing codebase comprehension, complex debugging, and end-to-end developer workflows. We achieved this by scaling training compute and diversifying the training environments.
Training Innovations
- Co-Training: Spark 1.2 was trained in tandem with Muse Code. This involved using rejection sampled harness trajectories and optimizing recipes for subagent coordination and goal compaction.
- Long-Horizon Capabilities: The model excels at massive tasks (e.g., generating entire repositories) by using:
- Planning for sequencing.
- Goal conditioning for focus.
- Context compaction for memory efficiency.
- Self-Improvement Loop: We used Spark 1.1 to create difficult instruction templates and environments. Spark 1.2 then graded the solutions, creating a high-quality, scalable dataset for its own training.
🔬 Case Study: GPU Kernel Optimization
To test the limits of the system, we tasked the agent with optimizing GPU kernels over a period of up to 24 hours, involving tool calls. The agent had to write, compile, and profile code to beat a baseline.
Constraints: The model was forbidden from using third-party libraries (like FLA) directly; it had to implement the logic in Triton using first-principles optimization.
Benchmark Results
| Kernel | Hardware | Strategy Used | Result |
|---|---|---|---|
| KDA | NVIDIA Hopper | Chunk-parallel prep + sequential inter-chunk scan; re-centering gated cumulative decay at chunk midpoint. | Substantial gain over FLA Triton baseline. |
| MLA | NVIDIA Hopper | Two-kernel Triton pipeline; kernel fusion and tiling; reused shared KV latent as both K and V. | Significant improvement over PyTorch reference. |
MLA Reference Parameters: The benchmark was conducted using the following LaTeX specifications:
🌐 Availability
Muse Spark 1.2 is accessible immediately via:
- The Muse Code terminal agent.
- The Meta Model API (with expanded global access).
We look forward to bringing you more powerful models and enhanced harness features soon!