Two Mac Studios, one enormous model, and no remote inference server in the middle. On October 15, Meta released Muse Spark Open, putting downloadable weights behind an unusually public collaboration with Apple.
The release targets coding and agents that carry out multi-step work. It is also a distribution play: developers can operate the model themselves instead of accessing it only through a provider’s service. Apple, meanwhile, gets a demanding showcase for its desktop hardware.
Muse Spark Open changes the access bargain
Meta lists 816 billion total parameters, with 42 billion active during inference, and a one-million-token context window. Its mixture-of-experts design activates only part of the model for each token. That reduces computation, not the need to store the wider model. Meta’s claims about capability remain the company’s own assessment, not a verdict from our testing.
This is a concrete change from the restricted-access Muse Spark launch NeonPulse covered in April. The new release makes weights available for local deployment, while retaining hosted options.
The desktop has a very large catch
Apple’s technical report explains the hardware bill in memory terms. Even compressed to the NVFP4 four-bit format, the weights occupy about 493 GB. Additional working memory means the demonstrated model does not fit on one 512 GB Mac Studio: computation must be distributed across multiple machines.
Apple reports roughly 25 generated tokens per second with two M5 Ultra Mac Studios using tensor parallelization, rising to about 37 with four. Those results used an 8,192-token context and measured 128 generated tokens. They are vendor benchmarks, not a guarantee for every agent or for the model’s maximum context.
MLX coordinates the workload, with low-latency communication over Thunderbolt helping the machines cooperate. This is local inference at workstation-cluster scale—not an announcement that an ordinary MacBook can run the same setup.
Downloadable does not mean unconditional
The published license and model card impose usage and attribution requirements, with separate terms for certain large model-as-a-service businesses. Meta also recommends safeguards around tool-using deployments. Open weights do not automatically make an agent safe or a commercial deployment unrestricted.
The strategic opening is bigger than a benchmark: developers gain another way to choose where their model runs. But the meter has not vanished; it has moved into hardware, electricity and administration. Apple and Meta are offering an alternative to renting every answer—and competing to own the infrastructure behind that choice.
