Meta has officially launched Muse Code, its first AI coding agent, powered by the Muse Spark 1.2 model, capable of handling complex software engineering tasks in the terminal. The model performs strongly on several benchmarks, trailing only Opus 5. Zuckerberg hinted at potential open-sourcing, sparking community interest.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Meta officially launches Muse Code, a terminal-based AI coding agent powered by Muse Spark 1.2, which trails only Opus 5 (max) on several benchmarks, sparking open-source discussions.
Meta officially released Muse Code beta on August 6, 2026, an AI coding agent that runs in the terminal, powered by the Muse Spark 1.2 model. Zuckerberg stated it can handle complex software engineering tasks in large codebases, including analyzing projects, planning modifications, writing code, running tools, and verifying results.
Muse Code employs a simple agent loop complemented by a set of asynchronous background agents that run continuously throughout the session, reducing redundant information gathering, autonomously executing subsequent steps, and deciding when to report back, thereby lowering latency and reducing reliance on human guidance.
After the product launch, netizens inquired about open-sourcing, and Zuckerberg responded, 'there will be more to share on this topic soon,' hinting at possible future open-sourcing.
Runtime design: Muse Code uses a local event log that records model calls, tool runs, approval actions, and code modifications as the single source of truth, enabling precise replay and safe recovery after restarts, even resuming from interruptions after crashes, suitable for long-horizon tasks.
Built-in skills: By default, it provides /plan (breaking tasks into plans requiring approval), /grill (repeatedly stress-testing plans), and /goal (persistently advancing toward goals until success), enhancing reliability and autonomy in task execution.
Muse Spark 1.2 is an upgraded coding capability version, focusing on improving code generation, debugging, codebase understanding, and end-to-end development. On Terminal-Bench 2.1, it trails only Opus 5 (max); on DeepSWE 1.1, it lags behind Opus 5 (max) and GPT-5.6 Terra (max); on Meta's internal benchmark, it trails only Opus 5 (max).
Training improvements include: co-training with Muse Code, incorporating rejection sampling trajectories and tool sets; large-scale training for long-horizon tasks using planning, goal constraints, and context compression; and a self-improvement mechanism that uses Muse Spark 1.1 to generate training data, enhancing instruction-following capabilities.
Muse Code pricing comes in standard and Contributor editions: standard edition input $1.25 per million tokens, cached input $0.15, output $4.25; Contributor edition input $0.10 per million tokens, cached input $0.002, output $0.20.
This move marks Meta's direct competition with OpenAI and Anthropic in the coding agent arena, and the product launch has sparked community discussions.
This article's information primarily comes from a report by Jiqizhixin, a secondary source, but it cites Meta's official blog and Zuckerberg's tweets. Benchmark data is as claimed by the source and has not been independently verified.
Meta has made a strong entry into the coding agent field with Muse Code and Muse Spark 1.2, with model performance approaching top-tier levels, and the possibility of open-sourcing has drawn attention, making future competition more intense.
Primary report
Primary source