Models
Muse Spark
Meta's first proprietary, closed-weight flagship model — a striking departure from the open-weight Llama line.
Model Facts
Provider
Meta Superintelligence Labs
Released
April 8, 2026
Context window
128K tokens
Weights
Closed
What it is
Muse Spark, launched April 8, 2026 by Meta's newly formed Superintelligence Labs, is Meta's first proprietary model — unlike every generation of Llama before it, Muse Spark's weights are not published for download. It marks a notable strategic reversal: Meta had spent years positioning Llama's open-weight releases as a point of differentiation from OpenAI, Anthropic, and Google, and Muse Spark's closed approach is widely read as a departure from that "open science" identity.
Why this matters
See Open Weight vs Closed Weight Models for the general trade-offs — Muse Spark is a live, high-profile example of a major lab moving in the closed direction after building its reputation on the opposite approach. Whether Meta continues developing Llama alongside Muse Spark, or Muse Spark gradually becomes the primary flagship while Llama's open releases slow down, is one of the more closely watched strategic questions in the industry right now.
Where this information comes from
Details here are based on Meta's own April 2026 announcement of Muse Spark. Meta has not published detailed third-party-verified benchmark comparisons for Muse Spark as of this writing — this entry is a seed example; update it via the admin dashboard as more concrete performance data becomes available.
Notable traits
- Meta's first closed-weight flagship, developed by the newly formed Superintelligence Labs group rather than the original Llama team.
- Positioned as a frontier-capability play, distinct from Llama's open, broadly-deployable positioning.
- Availability is through Meta's own products/API rather than downloadable weights.