By 
New Muse Spark 1.2 Model Powers Advanced Multi-Agent Software Engineering Across Massive Code Repositories
Meta has unveiled Muse Code, its first terminal-based artificial intelligence (AI) coding agent, marking a significant expansion of the company’s ambitions in developer tools and AI-assisted software engineering.
RELATED: Meta misses Q2 earnings forecast as AI spending, legal costs weigh on profit despite 28% revenue growth
The beta release, developed by Meta Superintelligence Labs (MSL) and powered by the newly introduced Muse Spark 1.2 model, is designed to automate complex software development tasks across large code repositories. The launch positions Meta in direct competition with AI coding assistants such as OpenAI Codex and Anthropic Claude Code.
Meta said both Muse Code and Muse Spark 1.2 are now available through the Meta Model API with expanded global access, representing another milestone in the company’s push toward frontier AI models.
Meta Expands AI Coding Portfolio
Announcing the launch, Meta Chief Executive Officer Mark Zuckerberg described Muse Code as the company’s next step in developing advanced AI systems capable of handling sophisticated engineering workflows.
“We’re excited to release Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our newest model. This marks our next step toward the frontier, with larger and much more capable models on the way,” Zuckerberg said.
The launch follows the release of Muse Spark 1.1 just a month earlier, highlighting Meta’s accelerated pace of AI product development.
Muse Code Automates Complex Software Engineering
According to Meta, Muse Code enables developers to complete end-to-end software engineering tasks using natural language instructions.
Rather than generating isolated code snippets, the AI agent is designed to plan software changes, write code, conduct testing and validate outputs across entire software repositories.
The platform employs a multi-agent architecture that coordinates several AI agents simultaneously, allowing complex engineering problems to be solved more quickly and accurately with minimal human intervention.
Persistent Background Agents Improve Productivity
One of Muse Code’s defining capabilities is its use of persistent asynchronous background agents.
Unlike conventional AI coding assistants that create and terminate individual agents for every request, Muse Code maintains specialised sub-agents throughout an entire development session.
These background agents continuously gather information, execute tasks and communicate with the primary coding agent when necessary, reducing redundant repository analysis and improving execution speed during complex, multi-stage projects.
According to Meta, this architecture significantly reduces latency while requiring less manual guidance from developers.
Restart-Safe Runtime Supports Long Projects
Meta also introduced a restart-safe runtime architecture that enables Muse Code to recover seamlessly after interruptions.
Every model interaction, tool execution, approval and code edit is recorded in a local event log, creating a complete operational history.
The company said this allows coding sessions to resume precisely where they stopped following system crashes or interruptions, making the platform suitable for long-running software engineering tasks involving thousands of tool interactions.
Muse Spark 1.2 Optimised for AI Software Development
Powering the new coding agent is Muse Spark 1.2, a coding-optimised large language model co-trained alongside Muse Code to maximise compatibility and performance.
Meta said the model delivers:
- Higher first-attempt coding success rates
- Cleaner tool execution
- Reduced need for repeated prompting
- Improved performance across long-running development projects
The model features a one-million-token context window, allowing it to process extremely large codebases and maintain contextual understanding across multiple files and extended engineering workflows.
Meta has priced API access at $1.25 per million input tokens and $4.25 per million output tokens.
Built for Long-Horizon Engineering Tasks
According to Meta, Muse Spark 1.2 was extensively trained on long-horizon coding assignments involving:
- Entire repository generation
- Large end-to-end software projects
- Automated research workflows
- Multi-file code generation
The model employs advanced planning mechanisms, goal conditioning and context compaction techniques to sustain performance throughout lengthy engineering processes while maintaining consistency across complex projects.
Self-Improving AI Training Framework
Meta disclosed that Muse Spark 1.2 benefited from a self-improvement training methodology.
Using the earlier Muse Spark 1.1 model, Meta generated complex coding environments and instruction-following scenarios, which were then evaluated automatically to identify the highest-quality solutions.
The resulting datasets were used to improve Muse Spark 1.2’s instruction-following capabilities and coding accuracy.
According to the company, this iterative learning process significantly enhanced the model’s ability to execute complex engineering instructions compared with its predecessor.
Case Study Demonstrates AI Kernel Optimisation
As part of its evaluation, Meta tested Muse Code’s ability to optimise GPU kernels over more than 1,000 tool calls during coding sessions lasting up to 24 hours.
Working within Muse Code’s agentic development environment, the AI system autonomously wrote, compiled, profiled and refined GPU kernel implementations for NVIDIA Hopper hardware.
The company reported that the platform consistently delivered substantial performance improvements over baseline implementations across Kernel Data Accelerator (KDA) and MLA kernel benchmarks.
Meta Strengthens Competition in AI Developer Tools
The launch of Muse Code reflects Meta’s growing investment in AI-powered software engineering as competition intensifies among leading AI companies to deliver increasingly autonomous coding assistants.
With persistent multi-agent capabilities, large-context reasoning, long-running workflow support and integrated software engineering automation, Muse Code positions Meta as a significant challenger in the rapidly evolving AI developer tools market.
The company indicated that Muse Spark 1.2 represents only the beginning of a broader roadmap, with more advanced frontier AI coding models expected in future releases.
