Meta Launches Muse Code Beta With Persistent Agents as Spark 1.2 Hits 82.9% on Terminal-Bench
Updated
Updated · InfoWorld · Aug 6
Meta Launches Muse Code Beta With Persistent Agents as Spark 1.2 Hits 82.9% on Terminal-Bench
3 articles · Updated · InfoWorld · Aug 6
Summary
Meta’s Muse Code beta for macOS and Linux uses persistent background agents that stay active across a session, tackling large codebases asynchronously instead of spinning up a new agent for each task.
A local event log records every model call, tool run, approval and edit, letting the system replay runs exactly, restart safely after crashes and cut repeated information gathering during multi-step work.
Meta said it co-trained Muse Spark 1.2 with Muse Code on longer assignments and broader development environments; the model scored 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE 1.1, trailing Claude Opus 5 on both.
Analysts said those benchmark gaps are hard to interpret because Meta tested each model with its preferred agent harness, while CIOs will care more about pass rates inside their own software pipelines.
Enterprise uptake may still be slow because persistent coding agents raise security, governance, cost and lock-in concerns; Meta’s Contributor tier costs $0.10 per million input tokens versus $1.25 for the standard tier but may be used to improve Meta’s products.