Meta ships Muse Spark Open, a 816B open-weight model

Meta built Muse Spark Open on top of its Muse Spark model and added safety measures, then released it as an open-weight model so organizations can run it on their own servers instead of routing through an external service. That matters because using an outside API brings cost and data-handling constraints, which the open-weight route avoids.

The 816B-parameter scale is a real obstacle to local use. The model needs about 1.6TB of memory at normal precision — too much for a typical single PC. Meta's answer was a collaboration with Apple: using MLX, Apple's AI development framework tuned for its chips, and quantization to cut the model down to about 493GB, then spreading the work across multiple Mac Studios.

Apple measured Muse Spark Open on two or four Mac Studios with M5 Ultra chips, connected over Thunderbolt 5 using RDMA for fast memory transfer between machines. On four Mac Studios, throughput reached about 36.7 tokens per second, faster than average human reading speed, according to Apple. The model is downloadable from Hugging Face and can run on NVIDIA and AMD environments too, and Meta signals it intends to keep giving developers free access to capable models.