Muse Code and Muse Spark 1.2
Meta's terminal coding agent Muse Code (beta) ships with Muse Spark 1.2, co-trained on harness trajectories.
Meta enters the Claude Code / Codex category. Muse Code runs persistent background sub-agents with an event log for crash-safe replay; Spark 1.2 was co-trained with the harness and used Spark 1.1 to generate training environments and grade solutions.
- Date
- Wednesday, 5 August 2026
- Lab
- Meta
- Kind
- product
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Coding benchmarks cited | Terminal-Bench 2.1, DeepSWE 1.1, Meta Internal Coding Bench scores shown only as charts in the post | company |
Spark 1.2 multimodal results from the Aug 20 post are ZeroBench 54.0% and CharXiv Reasoning 87.6%, per Meta. Meta later said Spark 1.2 weights would be opened, but they had not been released as of Sept 2 reporting.
Sources
- research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2
- ai.meta.com/blog/multimodal-intelligence-of-muse-spark-1-2
This record was checked against its sources on 6 October 2026. How we check