organisation · org/meta-superintelligence-labs

Meta Superintelligence Labs

Also called MSL, Meta AI, Meta

The company that made open weights a mainstream expectation stopped shipping them at the frontier. Meta's last frontier Llama release was Llama 4 Maverick on 5 April 2025. Its successor line arrived a year later: Muse Spark, introduced 8 April 2026 by Meta Superintelligence Labs, with the weights kept closed and the announcement saying only that Meta "hope[s] to open-source future versions of the model." In the OpenRouter catalog the change shows up as a change of prefix: meta-llama/ rows stop in 2025, meta/ rows begin in July 2026.

The catalogue's license line moved ahead of the weights themselves, and then stalled. Muse Spark 1.2's card on llm-releases, released 5 August 2026, was the first Spark flagship listed 'Open weights · Meta license (weights pending)' with weights 'Not released'source, accessed 2026-09-03; the card added that "Meta has signaled open weights are coming." One full flagship release later, on 2 September 2026, Muse Spark 1.3's card carries the same License and Weights lines, character for character. A month of "pending" spanning two releases, with nothing downloadable, is its own fact — and the catalogue contradicts itself about what the flag means: its own changelog still describes Glimmer as "its first open-weight model after the closed Muse Spark line." The only Spark-family weights actually shipped remain the distilled Glimmer 30B under Apache-2.0.

The licence went the other way. Llama shipped under Meta's own Llama 4 Community License Agreement, which required a separate licence from Meta for any product with more than 700 million monthly active users, obliged you to display "Built with Llama," and made you prefix derived model names with "Llama." The open-weight model Meta shipped on 10 August 2026, Muse Glimmer 30B, is Apache-2.0: no user threshold, no naming clause, and a licence text left unmodified. A usage policy still ships in the repository, but the Apache grant never references it, so it binds nothing — where the Llama agreement made adherence to its acceptable-use policy a condition of the licence. Meta's open weights got freer and smaller in the same move — Glimmer is a distillation of Muse Spark sized to run on one consumer GPU, not the flagship.

The third row is the one to read the terms on. Alongside meta/muse-spark-1.2 at $1.25per million tokensopenrouter-models, last checked 2026-09-14 input, Meta lists meta/muse-spark-1.2-contributor at $0.10per million tokensopenrouter-models, last checked 2026-09-14 — the same model, the same million-token window, with one condition stated in the listing: prompts and outputs may be used to improve Meta's productssource, accessed 2026-08-28. The gap between those two prices is a public number for what a prompt is worth to Meta, offered as a discount rather than asked for as a favour.

Facts

flagship weights
not released; Meta's 1.3 announcement (2026-09-02) names 'the Muse Spark open weights release' on its roadmapsource, accessed 2026-09-03
flagship weights listing
listed 'Open weights · Meta license (weights pending)' with weights 'Not released'source, accessed 2026-09-03
open weight release
Muse Glimmer 30B, Apache License 2.0source, accessed 2026-08-28
previous model line
Llama; last frontier release 2025-04-05 under the Llama 4 Community License Agreementsource, accessed 2026-08-28
contributor tier terms
prompts and outputs may be used to improve Meta's productssource, accessed 2026-08-28

Timeline

  1. Muse Spark 1.3 released; the flagship card's 'Open weights · Meta license (weights pending)' listing, unchanged since the 1.2 release, still shows weights not releasedsource
  2. Muse Glimmer 30B released under the Apache License 2.0source
  3. Muse Spark 1.2 releasedsource
  4. Muse Spark 1.1 launchedsource
  5. Muse Spark introduced as Meta Superintelligence Labs' first model, with closed weightssource
  6. Llama 4 Maverick released under the Llama 4 Community License Agreement — the last frontier Llama releasesource