Muse Spark Open on Mac Studio: What It Takes to Run Locally
Muse Spark Open can run locally on Mac hardware, but it needs at least two 512GB M5 Ultra Mac Studios. Here are the MLX, speed, license and privacy limits.
Direct answer: Muse Spark Open can run locally on Mac hardware, but not on an ordinary Mac or one Mac Studio. Meta and Apple describe a minimum local configuration of two Mac Studio computers, each using an M5 Ultra with 512GB of unified memory, connected for distributed inference with MLX. A four-machine setup is faster in Apple's specific benchmark. The downloadable model is open-weight under Meta's Muse Model License, not an unrestricted open-source release.
Meta released Muse Spark Open on October 15, 2026, while Apple published the distributed MLX work that makes the model practical on a small Mac Studio cluster. The release matters to developers because it combines downloadable weights, multimodal input and agentic coding capabilities with a desktop-scale alternative to data-center accelerators. It does not mean every Apple-silicon Mac can run the model.
What did Meta release?
Muse Spark Open is a mixture-of-experts model with 816 billion total parameters and about 42 billion active parameters for each generated token. Meta lists a context window of up to one million tokens and positions the model for coding, long-running agentic work, software engineering and multi-step reasoning.
The model accepts text, images, audio and video as input and generates text. “Multimodal” therefore does not mean it generates images, audio or video. Meta says the weights are available through Hugging Face and several cloud model catalogs, while hosted inference is available from named API providers.
Meta also says the release has additional safety post-training for standalone use. That is a release claim, not a guarantee that an application built on the model is safe, accurate or suitable for every workflow.
Can Muse Spark Open run locally on a Mac?
No, not in the configuration Apple evaluated. Apple says the model would need more than 1.6TB of memory for its BF16 parameters alone. The evaluated four-bit NVFP4 form reduces the model footprint to about 493GB, but inference also needs memory for intermediate work, so the whole system still does not fit on one 512GB Mac Studio.
The minimum documented local path splits the model across two Mac Studio computers with M5 Ultra and 512GB of unified memory per machine. Meta describes the pair as providing one terabyte of unified memory. Apple uses MLX 0.30.1 or later and the RDMA-over-Thunderbolt driver introduced in macOS 26.2; its setup connects the computers with two Thunderbolt cables for additional link bandwidth.
This is a specialist workstation cluster, not a checkbox feature on a consumer Mac. Readers comparing current desktop tiers can use the separate Mac mini versus Mac Studio buying guide, but Muse Spark Open's local requirements sit at the extreme end of that lineup.
How fast was Apple's benchmark?
At an 8,192-token context, Apple reports 15.74 generated tokens per second for two M5 Ultra Mac Studios using pipeline parallelism, 24.99 tokens per second for two machines using tensor parallelism and 36.73 tokens per second for four machines using tensor parallelism. Apple measured generation over 128 new tokens.
Those numbers are not a universal speed promise. They apply to Apple's named M5 Ultra hardware, NVFP4 model, MLX benchmark, interconnect and test lengths. Longer contexts change time-to-first-token and communication costs, and a real coding agent may spend time reading files, calling tools, compiling code or waiting on network services rather than only decoding text.
Four machines improve the measured decode rate, but they are not required merely to make the model fit. Two machines are the documented starting point; four improve parallel performance in Apple's test.
What does “local” mean here?
Local means the model weights and inference server can run on the Mac Studio cluster instead of sending every prompt to a hosted model API. Apple demonstrates an MLX server that exposes an OpenAI-compatible local endpoint for an agentic coding tool.
That does not make every workflow offline or private by default. An agent can still browse the web, call remote tools, install packages, upload telemetry or access files according to the surrounding application's permissions. Local inference changes where the model computation runs; it does not automatically audit the agent, its extensions or its data handling. Mac users evaluating agent permissions can also read the macOS Full Disk Access and AI agent privacy guide.
Is Muse Spark Open open source?
Meta calls the release open-weight. The weights can be downloaded, but Meta releases them under the Muse Model License with limited commercial terms. Open-weight access is not the same as an unrestricted open-source license, and developers should review the current license for their intended use rather than relying on the word “open.”
The release supports several inference routes, including MLX, SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers and llama.cpp. Hardware and quantization requirements vary by route. Apple's results establish one reviewed Mac configuration, not compatibility or equal performance for every engine, Mac model or checkpoint format.
Who should use the Mac Studio route?
The local cluster makes the most sense for research teams and developers who specifically need control of the weights, local experimentation, large-context agentic coding or a Mac-native distributed environment and can justify multiple top-memory Mac Studios. It is not the practical default for a developer who only wants to try the model.
For evaluation, Meta lists hosted API providers and cloud model catalogs that avoid the local hardware requirement. For single-Mac local work, Apple's report also benchmarks smaller models that fit on one 512GB Mac Studio. Those are different model choices, not reduced-memory modes for the full Muse Spark Open model.
What should developers check before starting?
- Confirm whether the goal requires local weights or whether a hosted API is sufficient.
- For the reviewed Mac route, budget for at least two M5 Ultra Mac Studios with 512GB unified memory each.
- Check the required macOS, MLX and Thunderbolt RDMA setup against Apple's current documentation.
- Choose pipeline or tensor parallelism based on fit, latency and communication needs rather than assuming four machines are mandatory.
- Review the Muse Model License and the application's security, file-access, tool and network behavior.
- Treat Apple's benchmark as configuration-specific evidence, not a promise for a different context, application or hardware topology.
Muse Spark Open has been released, and Apple has documented a credible local route. The useful takeaway is narrower than the headline: this is desktop-cluster AI for specialized teams, not a frontier model that suddenly runs on any Mac.