What the reports cover
Artificial Analysis’s September 2 report evaluates Muse Spark 1.3, including separate effort configurations and a limited preview. Its August 10 Glimmer report examines a smaller open-weight offering. Meta’s own August 10 announcement confirms Glimmer’s downloadable weights under Apache 2.0. These are different models, with different access routes.
The everyday takeaway
When a newer model appears, older results remain useful context, but their rankings age. Compare the same task and configuration, and check the actual license before planning a local installation. A hosted assistant and a downloadable checkpoint can have very different requirements.
Our coverage boundary
We have not run these models. The benchmark discussion describes the evaluator’s published work, not our measurements. These are historical releases; our separate Muse Wonder briefing covers Meta’s October 12 announcement and an outside evaluation.