Meta's first closed model——
the open-AI flag comes down
Muse Spark is Meta's first fully closed-source model, available only to select partners in preview. With no open-source commitment attached, it marks a meaningful departure from the open-weight identity that Llama built over years.
Llama's openness was
a competitive strategy too
Meta positioned Llama's open-weight releases as "democratising AI" — and it was that, but it was also a calculated strategy against OpenAI and Anthropic's closed models. An open ecosystem around Meta's models built developer loyalty and created competitive pressure on proprietary alternatives. Walking that back, even partially, carries real signal.
Muse Spark is Meta's first fully closed-source model, currently in a select-partner preview. Future open-sourcing is described only as "under consideration" — no commitment has been made.
Hardware integration and
competitive lead protection
Two motivations plausibly explain the closed-source decision: tight hardware integration with Meta's glasses, and protecting a first-mover advantage in physical AI.
Whether this is a one-off decision or the start of a broader shift is not yet clear. The Llama series continues open-weight — so the current state is "Meta now does both." If Muse Spark succeeds, the incentive to apply the same model to future frontier releases will grow.
Projects banking on Meta's
open posture should take note
Llama itself remains open-weight and existing projects are not affected. But organisations whose long-term strategy assumes Meta will keep releasing frontier models as open weights should treat this as a signal worth tracking. If future leading Meta models follow Muse Spark's pattern, the assumption of free access to frontier-level Meta AI will need revisiting. The cost of monitoring is low; the cost of being caught off guard is higher.