Muse Spark

weights

closed

data

closed

code

closed

license

Proprietary(meta.ai/Meta AI app + private API preview

No weights released - a closed assistant-surface model offered only through a private API preview. Meta's newsroom post says Muse Spark powers the Meta AI app and website and will be offered in private preview via API to select partners, and names no weights, corpus or training code anywhere on it; Meta's engineering blog is not publicly readable.

• https://ai.meta.com/blog/introducing-muse-spark-msl/ recorded 2026-06-04

  Meta's engineering blog for Muse Spark. It answers HTTP 400 to a plain fetch.

• https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs/ recorded 2026-08-13

  Meta primary newsroom post - "Muse Spark is our most powerful model yet. It currently powers the Meta AI app and website, and will be rolling out to WhatsApp, Instagram, Facebook, Messenger, and AI glasses"; "We will also be offering the model in private preview via API to select partners". No weights, corpus or training code.

Powers the Meta AI assistant on meta.ai and in the Meta AI app, and across WhatsApp, Instagram, Facebook, Messenger and the Ray-Ban and Oakley glasses - an enormous consumer surface - but Meta discloses no Muse Spark-specific user count. Placed at 4 on the consumer-app surface it powers, attributed honestly to that surface rather than to the model. Not 5 because there is no standalone figure behind it.

• https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs/ recorded 2026-08-13

  "It currently powers the Meta AI app and website, and will be rolling out to WhatsApp, Instagram, Facebook, Messenger, and AI glasses"; 2026-05-12 update - "When we launched Muse Spark last month, we said we'd bring it everywhere. Today, we are". No Muse Spark user count published.

• https://ai.meta.com/blog/introducing-muse-spark-msl/ recorded 2026-06-04

  Meta's engineering blog for Muse Spark. It answers HTTP 400 to a plain fetch.

The 5 rests on a top-five arena placement and a rank of 14 out of 188 on an independent index, both re-derivable. The arena.ai text board - lmarena.ai redirects to it - ranks muse-spark-1.2 (xHigh) 4th of 391 models at 1498 +/-10, behind only three Anthropic entries, and Artificial Analysis corroborates independently at 57 on its Intelligence Index, rank 14 of 188. The placement belongs to the 1.2 xHigh release rather than to the tier generally, which is what the recorded value says. No HLE or FrontierScience figures are carried, because Meta's own site is not publicly readable and no other source stands behind them.

• https://arena.ai/leaderboard/text recorded 2026-08-14

  arena.ai text leaderboard (lmarena.ai 301s here) - rank 4 "muse-spark-1.2 (xHigh)", Meta, Proprietary, score 1498 +/-10 over 3,280 votes, behind claude-fable-5 1506, claude-opus-4-6-high 1505 and claude-opus-4-7-high 1502, out of 391 models and 7,779,985 votes

• https://artificialanalysis.ai/models/muse-spark-1-2 recorded 2026-08-14

  "scores 57 on the Artificial Analysis Intelligence Index, placing it well above average among comparable models (median - 34)"; Intelligence rank "# 14 / 188". No deprecation banner.

• https://ai.meta.com/blog/introducing-muse-spark-msl/ recorded 2026-06-04

  The former source of the HLE 58% and FrontierScience 38% figures. It answers HTTP 400 to the repo's fetcher, so those figures remain unsourced and have been dropped from the value.