Meta has introduced Muse Code (beta), a terminal-based AI coding agent powered by its latest Muse Spark 1.2 model. Designed for complex software engineering tasks across large code repositories, the tool can plan code changes, write code, and validate results with minimal user intervention. The company said Muse Code marks its next step toward more capable AI models, with larger models planned for future releases.
Muse Code features
Muse Code uses a simple agent loop alongside multiple persistent asynchronous background agents that work with the main coding agent to handle complex, multi-step software engineering tasks. Unlike task-specific agents, these remain active throughout a session to:
- Reduce redundant information gathering
- Lower latency
- Minimize the need for manual steering
- Carry out subsequent steps independently
- Communicate updates back to the main agent when needed
The tool also maintains a local event log that records every:
- Model call
- Tool run
- Approval
- Code edit
Meta said this serves as a single source of truth, making the runtime replay-exact and restart-safe so sessions can resume from where they stopped after interruptions, allowing long-running tasks to continue without losing progress.
Muse Code also includes three built-in commands:
- /plan – Creates an approval-gated execution plan
- /grill – Reviews and stress-tests the plan
- /goal – Works toward completing a specified objective
Muse Spark 1.2
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1 with improvements in:
- Code generation
- Complex debugging
- Codebase understanding
- End-to-end software engineering workflows
- General AI agent capabilities
Meta said it significantly increased training compute for coding tasks while expanding the diversity of training environments.
The model was co-trained with Muse Code using:
- Rejection-sampled harness trajectories
- Recipe optimizations for goals
- Context compaction
- Subagents
- Integration of the Muse Code toolset to maximize harness compatibility
Meta said Muse Spark 1.2 was also trained on long-horizon software engineering tasks, including:
- Whole-repository code generation
- Large end-to-end projects
- Automated research
It uses planning to sequence work, goal conditioning to maintain direction and context compaction to retain the information needed during lengthy development sessions.
In addition, Meta used Muse Spark 1.1 to generate challenging coding environments and instruction-following templates. The earlier model evaluated candidate solutions based on how well they satisfied the required objectives, producing a scalable training dataset for Muse Spark 1.2.
According to the company, this self-improvement process enables the model to follow complex coding instructions more accurately than its predecessor.
Kernel optimization testing
Meta said it evaluated Muse Spark 1.2 on GPU kernel optimization tasks involving more than 1,000 tool calls over periods of up to 24 hours. Using Muse Code, the model repeatedly wrote, compiled, profiled and progressively optimized KDA and MLA kernels for NVIDIA Hopper GPUs, improving performance relative to the provided baseline implementation.
Pricing and Availability
Meta said Muse Code (beta) is available for macOS and Linux. Users can install the tool using:
curl -fsSL https://dev.meta.ai/install.sh | bashMuse Spark 1.2 is also available through Muse Code and the Meta Model API with expanded global access. The model is offered in two variants:
| Model | Context window | Input | Cached input | Output |
|---|---|---|---|---|
| muse-spark-1.2-contributor | 1M tokens | $0.10/M tokens | $0.002/M tokens | $0.20/M tokens |
| muse-spark-1.2 | 1M tokens | $1.25/M tokens | $0.15/M tokens | $4.25/M tokens |
Meta said the muse-spark-1.2-contributor variant is used to improve its products, while muse-spark-1.2 is not used for product improvement. The company also plans to introduce additional harness features and more powerful AI models.