№ 0101·Resource★ Pick

Muse Spark 1.3 tutorial: testing Meta's efficiency claims

DataCamp's Josep Ferrer ran Muse Spark 1.3 on three real coding tasks. Two used 23–32% fewer completion tokens, but a refactor used 70% more, for a net 12% cost increase.

He also notes that Muse Code defaults to the muse-spark-1.3-contributor model, which lets Meta use code for product improvement, and that the documented 'ultra' reasoning level silently falls back to xhigh. Uncached input fell 19–22%, short of Meta's 25% claim.