Model and agent are related, not identical

Muse Spark 1.2 is Meta’s coding-focused model update. Muse Code is the terminal coding agent powered by that model. Meta says they were co-trained so the model would work effectively with the agent’s tools, goals, compaction, and subagent patterns.

Model
Muse Spark 1.2
Coding-focused update
Agent
Muse Code
Terminal beta
Training relationship
Co-trained
Current Meta statement
Task horizon
Long-running
Repository-scale emphasis
Verified system mapConceptual · not to scale
The sequence groups concepts Meta documents; it is not a complete internal architecture diagram.

What Meta currently documents

AreaCurrent official statementGuide boundary
Coding focusMore training compute and environment diversity for codingNo invented universal benchmark winner
Co-trainingHarness trajectories, goals, compaction, subagents, and toolsNo claim that model and product are interchangeable
Long horizonWhole-repository generation and end-to-end projectsNo guarantee for every repository or duration
AvailabilityMuse Code and Meta Model API with expanded global accessNo inferred country list

Meta also publishes benchmark and kernel-optimization material. Those results explain how Meta evaluated the model, but they do not settle which coding product is best for every repository, language, policy, or budget.

Why the runtime matters

A coding model does not operate in isolation. Muse Code adds tools, a main loop, persistent background agents, validation, approvals, edits, and an event history. That surrounding system determines what context the model receives, which actions it can take, and how work survives an interruption.

Read parallel subagents for the documented background-agent pattern and event log and resume for the restart-safe runtime claim. The getting started guide brings those pieces back to a safe first project.