Meta Muse Spark Open: Open-Weight AI Model Needs 1TB RAM

What happened

Meta has released Muse Spark Open, an open-weight AI model that the company describes as "the most capable US open-weight model to date." According to Engadget, the model can be run on Macs, but only with a very large memory requirement of around 1TB of RAM.

The reporting available to us covers the headline claims and the hardware constraint. Release timing, licensing terms and independent benchmark results are not set out in the material we reviewed, so readers should treat the capability claim as the company's own assessment for now.

Why it matters

Open-weight releases matter because they allow organisations to inspect, adapt and run a model outside a single provider's hosted service. That flexibility is attractive for teams with data-residency, privacy or customisation requirements. For leaders deciding where AI workloads should live, however, the hardware bar is the central fact: a capable model only changes how an organisation operates if the infrastructure to serve it is affordable and manageable.

For digital transformation programmes, this points to a trade-off between control and operational burden. Running a model locally or privately can bring data and response times closer to the business, but it moves responsibility for procurement, maintenance, security and governance onto the organisation itself.

The Renascence take

Most coverage will focus on the ranking claim. The more useful question is who can actually use the model, and what it costs them in money, time and dependency.

The headline figure that matters is the 1TB memory requirement, not the benchmark boast, and that is where the experience gap sits. A model only a handful of well-resourced teams can run creates a service-design problem: the people who will live with its answers are often furthest from the hardware decisions. Customer-obsessed operators should test smaller, task-specific models against real service journeys first, and treat the largest open model as a capability to reach for selectively rather than a default.