Meta this week released Muse Spark Open, an open-weight version of its flagship Muse Spark artificial intelligence model, following through on a pledge chief executive Mark Zuckerberg made in a September post that an open-weights release was "coming soon."
Muse Spark launched in April as the first public model from Meta's Superintelligence Labs, but unlike the company's earlier Llama models it shipped closed, available only inside Meta's own products and through a private-preview API for select partners. That broke sharply with Meta's longstanding pitch that open models were a competitive and strategic advantage over closed rivals such as OpenAI, Anthropic and Google, and it drew criticism from developers who had built tools around freely downloadable Llama weights.
Meta began walking the closed approach back over the summer. In August it released Muse Glimmer, a smaller, 30-billion-parameter model distilled for local use and licensed under Apache 2.0, while continuing to keep Muse Spark itself closed. Zuckerberg's September commitment to open Muse Spark's weights, and this week's release of Muse Spark Open, extends that same openly licensed approach to the larger, more capable model line for the first time.
Meta has not published full technical documentation or license terms alongside the release, and independent benchmark trackers have only tested earlier, closed versions of Muse Spark. According to reporting on Meta's recent Muse Spark updates, the benchmarking firm Artificial Analysis has previously rated the model family competitive with offerings from OpenAI and xAI, though Meta's own performance claims for its newest releases have generally outpaced what outside testers have independently confirmed.
Like its closed predecessor, Muse Spark is described by Meta as a multimodal, reasoning-capable model built for tasks such as coding, long-running agent workflows and handling ambiguous instructions by asking clarifying questions rather than guessing. Meta has said successive Muse Spark updates have improved on those fronts, though the company has not detailed how Muse Spark Open's capabilities compare to the hosted, closed version still running inside Meta's own apps.
Reaction from the developer community has been mixed. Meta's reversal from closed to open licensing within months, after its launch-day break from the open Llama approach, has left some engineers skeptical that the company's open-source commitments will hold for future releases, even as others welcomed the ability to self-host a frontier-class Meta model for the first time this year.
The open release matters for a developer community that has grown wary of Meta's shifting stance: it lets outside engineers download and run Muse Spark on their own infrastructure rather than relying solely on Meta's hosted API, a capability closed competitors generally do not offer. Whether Meta continues releasing future Muse Spark versions openly, or reverts to closed launches as it did in April, remains an open question.