Meta’s Muse Spark Open brings an 816B model to Macs and major runtimes

What changed · following since Oct 16 Oct 17 — Apple reports Muse Spark Open exceeding 36 tokens per second on a four-Mac-Studio M5 Ultra cluster at 8,192-token context.

Muse Spark Open is best understood as a flexible deployment target for private, long-context agentic coding rather than a conventional local desktop model. Its open weights, multimodal design and broad runtime and hosting coverage give teams multiple operating paths, while the memory and hardware requirements keep production evaluation essential.

• Muse Spark Open is a sparse mixture-of-experts model with 816B total parameters, 42B active per token and a 1M-token context window; it supports native image and video perception plus audio input.
• Meta released the model weights on October 15, 2026, with access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, plus deployment through Amazon SageMaker JumpStart and downloads through Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models.
• Local execution is designed for two Apple M5 Ultra Mac Studios; TechCrunch reported a requirement of at least 1TB of RAM, making the desktop path a specialized deployment rather than an ordinary developer workstation.
• vLLM provided day-zero support for NVIDIA and AMD serving, initial TPU support, tensor and pipeline parallelism, and speculative decoding through DSpark; Meta also lists SGLang, TensorRT-LLM, Transformers, llama.cpp and Unsloth as supported runtimes.
• Meta paired Muse Spark Open with Muse Code for agentic workflows, including coding fixes, new features and a full build for a Mastodon client, as well as work with DaVinci Resolve.
• Caveat: Apple reports more than 36 tokens per second at an 8,192-token context on a cluster of four M5 Ultra Mac Studios using MLX and RDMA; this is a specialized hardware benchmark, not a typical single-workstation result.

Bottom line Muse Spark Open is worth testing for private, long-context agentic workloads, but its practical advantage depends on whether your serving environment can absorb its exceptional memory and hardware footprint.