Meta Debuts Muse Code Beta and Muse Spark 1.2 for Agentic Engineering
The terminal-based agent leverages persistent background processes and a 1-million-token window to manage large-scale repositories.
Meta has released Muse Code in beta, a terminal-based coding agent powered by the company's latest coding-centric model, Muse Spark 1.2. The system marks a shift toward autonomous software engineering by coordinating persistent, asynchronous background agents to handle complex tasks across expansive codebases.
According to Meta Research, Muse Code utilizes async background agents that remain active throughout a session. This architecture reduces latency and eliminates the need for redundant information gathering during long-horizon tasks. To ensure reliability, the runtime maintains a local event log of every model call and tool execution, which Meta states makes the agent both restart-safe and replay-exact.
Technical Specifications and Performance
The underlying model, Muse Spark 1.2, is built for scale, featuring a context window of 1,048,576 tokens. This capacity allows the agent to ingest and analyze vast amounts of repository data simultaneously. In terms of performance, Muse Code achieved a score of 82.9 on the Terminal-Bench 2.1 benchmark. Additionally, data from the BenchAlign leaderboard places Muse Spark 1.2 at rank #49 out of 216 models, with a score of 60.25/100.
For developers accessing the model via API, pricing is set at $1.25 per million input tokens and $4.25 per million output tokens, according to data from OpenRouter and Artificial Analysis.
The Shift to Agentic Coding
This release signals Meta's strategic move beyond simple autocomplete or chat-based code generation. By focusing on "agentic" coding, Meta is targeting autonomous capabilities capable of managing whole repositories and executing long-running operations. The goal is to enable the system to perform high-complexity engineering work—such as GPU kernel optimization—that may require over 1,000 individual tool calls to complete.
Industry Implications
The introduction of Muse Code places Meta in direct competition with other agentic coding tools that aim to reduce the manual overhead of debugging and refactoring. By integrating a massive context window with a state-aware runtime, Meta is attempting to solve the "forgetfulness" and instability often associated with LLM-based agents in large-scale production environments.
What to Watch
As Muse Code remains in beta, the industry will be watching for real-world deployment data regarding its stability in diverse enterprise environments. While benchmark scores like Terminal-Bench provide a baseline, the true test will be the agent's ability to maintain coherence across thousands of tool calls in non-synthetic, legacy codebases.