Meta's first non-open-weights frontier-class model scores 52 on the index with strong token efficiency but unremarkable agentic results, and has no API yet.
Key takeaways · AI-distilled
Muse Spark is Meta's first frontier-class release since Llama 4 Maverick in April 2025 and its first model not released as open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition →. Artificial Analysis benchmarked it with early access at 52, behind only Gemini 3.1 Pro, GPT-5.4 and Claude Opus 4.6.
It used 58M output tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → to run the Intelligence Index, about the same as Gemini 3.1 Pro Preview (57M) and far fewer than Claude Opus 4.6 at max effort (157M) or GPT-5.4 xhigh (120M).
It is the second-most capable vision model Artificial Analysis has benchmarked, scoring 80.5% on MMMU-Pro, behind only Gemini 3.1 Pro Preview (82.4%).
Agentic performance does not stand out: on GDPval-AA it scores 1427, behind Claude Sonnet 4.6 (1648) and GPT-5.4 (1676) but ahead of Gemini 3.1 Pro Preview (1320).
At publication there was no public API, though Meta said one was coming; the model is being integrated into Meta AI, Facebook, Instagram and Threads.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
